Skip to content

LLM Connectivity Reference

Here we handle connections to various LLM services, proprietary and open source.

Handle connections to different LLM providers.

AnthropicConversation

Bases: Conversation

Conversation class for the Anthropic model.

Source code in biochatter/llm_connect/anthropic.py
class AnthropicConversation(Conversation):
    """Conversation class for the Anthropic model."""

    def __init__(
        self,
        model_name: str,
        prompts: dict,
        correct: bool = False,
        split_correction: bool = False,
    ) -> None:
        """Connect to Anthropic's API and set up a conversation with the user.

        Also initialise a second conversational agent to provide corrections to
        the model output, if necessary.

        Args:
        ----
            model_name (str): The name of the model to use.

            prompts (dict): A dictionary of prompts to use for the conversation.

            split_correction (bool): Whether to correct the model output by
                splitting the output into sentences and correcting each
                sentence individually.

        """
        super().__init__(
            model_name=model_name,
            prompts=prompts,
            correct=correct,
            split_correction=split_correction,
        )

        self.ca_model_name = "claude-3-5-sonnet-20240620"
        # TODO make accessible by drop-down

    def set_api_key(self, api_key: str, user: str | None = None) -> bool:
        """Set the API key for the Anthropic API.

        If the key is valid, initialise the conversational agent. Optionally set
        the user for usage statistics.

        Args:
        ----
            api_key (str): The API key for the Anthropic API.

            user (str, optional): The user for usage statistics. If provided and
                equals "community", will track usage stats.

        Returns:
        -------
            bool: True if the API key is valid, False otherwise.

        """
        client = anthropic.Anthropic(
            api_key=api_key,
        )
        self.user = user

        try:
            client.count_tokens("Test connection")
            self.chat = ChatAnthropic(
                model_name=self.model_name,
                temperature=0,
                api_key=api_key,
            )
            self.ca_chat = ChatAnthropic(
                model_name=self.ca_model_name,
                temperature=0,
                api_key=api_key,
            )
            if user == "community":
                self.usage_stats = get_stats(user=user)

            return True

        except anthropic._exceptions.AuthenticationError:
            self._chat = None
            self._ca_chat = None
            return False

    def _primary_query(self, **kwargs) -> tuple:
        """Query the Anthropic API with the user's message.

        Return the response using the message history (flattery system messages,
        prior conversation) as context. Correct the response if necessary.

        Args:
        ----
            **kwargs: Keyword arguments (not used by this basic Anthropic implementation,
                     but accepted for compatibility with the base Conversation interface)

        Returns:
        -------
            tuple: A tuple containing the response from the Anthropic API and
                the token usage.

        """
        if kwargs:
            warnings.warn(f"Warning: {kwargs} are not used by this class", UserWarning)

        try:
            history = self._create_history()
            response = self.chat.generate([history])
        except (
            anthropic._exceptions.APIError,
            anthropic._exceptions.AnthropicError,
            anthropic._exceptions.ConflictError,
            anthropic._exceptions.NotFoundError,
            anthropic._exceptions.APIStatusError,
            anthropic._exceptions.RateLimitError,
            anthropic._exceptions.APITimeoutError,
            anthropic._exceptions.BadRequestError,
            anthropic._exceptions.APIConnectionError,
            anthropic._exceptions.AuthenticationError,
            anthropic._exceptions.InternalServerError,
            anthropic._exceptions.PermissionDeniedError,
            anthropic._exceptions.UnprocessableEntityError,
            anthropic._exceptions.APIResponseValidationError,
        ) as e:
            return str(e), None

        msg = response.generations[0][0].text
        token_usage_raw = response.llm_output.get("token_usage")
        token_usage = self._extract_total_tokens(token_usage_raw)

        self.append_ai_message(msg)

        return msg, token_usage

    def _create_history(self) -> list:
        """Create a history of messages for the Anthropic API.

        Returns
        -------
            list: A list of messages, with the last message being the most
                recent.

        """
        history = []
        # extract text components from message contents
        msg_texts = [m.content[0]["text"] if isinstance(m.content, list) else m.content for m in self.messages]

        # check if last message is an image message
        is_image_message = False
        if isinstance(self.messages[-1].content, list):
            is_image_message = self.messages[-1].content[1]["type"] == "image_url"

        # find location of last AI message (if any)
        last_ai_message = None
        for i, m in enumerate(self.messages):
            if isinstance(m, AIMessage):
                last_ai_message = i

        # Aggregate system messages into one message at the beginning
        system_messages = [m.content for m in self.messages if isinstance(m, SystemMessage)]
        if system_messages:
            history.append(
                SystemMessage(content="\n".join(system_messages)),
            )

        # concatenate all messages before the last AI message into one message
        if last_ai_message is not None:
            history.append(
                HumanMessage(
                    content="\n".join([m for m in msg_texts[:last_ai_message]]),
                ),
            )
            # then append the last AI message
            history.append(
                AIMessage(
                    content=msg_texts[last_ai_message],
                ),
            )

            # then concatenate all messages after that
            # into one HumanMessage
            history.append(
                HumanMessage(
                    content="\n".join(
                        [m for m in msg_texts[last_ai_message + 1 :]],
                    ),
                ),
            )

        # else add human message to history (without system messages)
        else:
            last_system_message = None
            for i, m in enumerate(self.messages):
                if isinstance(m, SystemMessage):
                    last_system_message = i
            history.append(
                HumanMessage(
                    content="\n".join(
                        [m for m in msg_texts[last_system_message + 1 :]],
                    ),
                ),
            )

        # if the last message is an image message, add the image to the history
        if is_image_message:
            history[-1].content = [
                {"type": "text", "text": history[-1].content},
                {
                    "type": "image_url",
                    "image_url": {
                        "url": self.messages[-1].content[1]["image_url"]["url"],
                    },
                },
            ]
        return history

    def _correct_response(self, msg: str) -> str:
        """Correct the response from the Anthropic API.

        Send the response to a secondary language model. Optionally split the
        response into single sentences and correct each sentence individually.
        Update usage stats.

        Args:
        ----
            msg (str): The response from the Anthropic API.

        Returns:
        -------
            str: The corrected response (or OK if no correction necessary).

        """
        ca_messages = self.ca_messages.copy()
        ca_messages.append(
            HumanMessage(
                content=msg,
            ),
        )
        ca_messages.append(
            SystemMessage(
                content="If there is nothing to correct, please respond with just 'OK', and nothing else!",
            ),
        )

        response = self.ca_chat.generate([ca_messages])

        correction = response.generations[0][0].text
        token_usage = response.llm_output.get("token_usage")

        return correction

__init__(model_name, prompts, correct=False, split_correction=False)

Connect to Anthropic's API and set up a conversation with the user.

Also initialise a second conversational agent to provide corrections to the model output, if necessary.


model_name (str): The name of the model to use.

prompts (dict): A dictionary of prompts to use for the conversation.

split_correction (bool): Whether to correct the model output by
    splitting the output into sentences and correcting each
    sentence individually.
Source code in biochatter/llm_connect/anthropic.py
def __init__(
    self,
    model_name: str,
    prompts: dict,
    correct: bool = False,
    split_correction: bool = False,
) -> None:
    """Connect to Anthropic's API and set up a conversation with the user.

    Also initialise a second conversational agent to provide corrections to
    the model output, if necessary.

    Args:
    ----
        model_name (str): The name of the model to use.

        prompts (dict): A dictionary of prompts to use for the conversation.

        split_correction (bool): Whether to correct the model output by
            splitting the output into sentences and correcting each
            sentence individually.

    """
    super().__init__(
        model_name=model_name,
        prompts=prompts,
        correct=correct,
        split_correction=split_correction,
    )

    self.ca_model_name = "claude-3-5-sonnet-20240620"

set_api_key(api_key, user=None)

Set the API key for the Anthropic API.

If the key is valid, initialise the conversational agent. Optionally set the user for usage statistics.


api_key (str): The API key for the Anthropic API.

user (str, optional): The user for usage statistics. If provided and
    equals "community", will track usage stats.

bool: True if the API key is valid, False otherwise.
Source code in biochatter/llm_connect/anthropic.py
def set_api_key(self, api_key: str, user: str | None = None) -> bool:
    """Set the API key for the Anthropic API.

    If the key is valid, initialise the conversational agent. Optionally set
    the user for usage statistics.

    Args:
    ----
        api_key (str): The API key for the Anthropic API.

        user (str, optional): The user for usage statistics. If provided and
            equals "community", will track usage stats.

    Returns:
    -------
        bool: True if the API key is valid, False otherwise.

    """
    client = anthropic.Anthropic(
        api_key=api_key,
    )
    self.user = user

    try:
        client.count_tokens("Test connection")
        self.chat = ChatAnthropic(
            model_name=self.model_name,
            temperature=0,
            api_key=api_key,
        )
        self.ca_chat = ChatAnthropic(
            model_name=self.ca_model_name,
            temperature=0,
            api_key=api_key,
        )
        if user == "community":
            self.usage_stats = get_stats(user=user)

        return True

    except anthropic._exceptions.AuthenticationError:
        self._chat = None
        self._ca_chat = None
        return False

AzureGptConversation

Bases: GptConversation

Conversation class for the Azure GPT model.

Source code in biochatter/llm_connect/azure.py
class AzureGptConversation(GptConversation):
    """Conversation class for the Azure GPT model."""

    def __init__(
        self,
        deployment_name: str,
        model_name: str,
        prompts: dict,
        correct: bool = False,
        split_correction: bool = False,
        version: str | None = None,
        base_url: str | None = None,
        update_token_usage: Callable | None = None,
    ) -> None:
        """Connect to Azure's GPT API and set up a conversation with the user.

        Extends GptConversation.

        Args:
        ----
            deployment_name (str): The name of the Azure deployment to use.

            model_name (str): The name of the model to use. This is distinct
                from the deployment name.

            prompts (dict): A dictionary of prompts to use for the conversation.

            correct (bool): Whether to correct the model output.

            split_correction (bool): Whether to correct the model output by
                splitting the output into sentences and correcting each
                sentence individually.

            version (str): The version of the Azure API to use.

            base_url (str): The base URL of the Azure API to use.

            update_token_usage (Callable): A function to update the token usage
                statistics.

        """
        super().__init__(
            model_name=model_name,
            prompts=prompts,
            correct=correct,
            split_correction=split_correction,
            update_token_usage=update_token_usage,
        )

        self.version = version
        self.base_url = base_url
        self.deployment_name = deployment_name

    def set_api_key(self, api_key: str, user: str | None = None) -> bool:
        """Set the API key for the Azure API.

        If the key is valid, initialise the conversational agent. No user stats
        on Azure.

        Args:
        ----
            api_key (str): The API key for the Azure API.

            user (str, optional): The user for usage statistics.

        Returns:
        -------
            bool: True if the API key is valid, False otherwise.

        """
        try:
            self.chat = AzureChatOpenAI(
                deployment_name=self.deployment_name,
                model_name=self.model_name,
                openai_api_version=self.version,
                azure_endpoint=self.base_url,
                openai_api_key=api_key,
                temperature=0,
            )
            self.ca_chat = AzureChatOpenAI(
                deployment_name=self.deployment_name,
                model_name=self.model_name,
                openai_api_version=self.version,
                azure_endpoint=self.base_url,
                openai_api_key=api_key,
                temperature=0,
            )

            self.chat.generate([[HumanMessage(content="Hello")]])
            self.user = user if user is not None else "Azure Community"

            return True

        except openai._exceptions.AuthenticationError:
            self._chat = None
            self._ca_chat = None
            return False

    def _update_usage_stats(self, model: str, token_usage: dict) -> None:
        if self._update_token_usage is not None:
            self._update_token_usage(self.user, model, token_usage)

__init__(deployment_name, model_name, prompts, correct=False, split_correction=False, version=None, base_url=None, update_token_usage=None)

Connect to Azure's GPT API and set up a conversation with the user.

Extends GptConversation.


deployment_name (str): The name of the Azure deployment to use.

model_name (str): The name of the model to use. This is distinct
    from the deployment name.

prompts (dict): A dictionary of prompts to use for the conversation.

correct (bool): Whether to correct the model output.

split_correction (bool): Whether to correct the model output by
    splitting the output into sentences and correcting each
    sentence individually.

version (str): The version of the Azure API to use.

base_url (str): The base URL of the Azure API to use.

update_token_usage (Callable): A function to update the token usage
    statistics.
Source code in biochatter/llm_connect/azure.py
def __init__(
    self,
    deployment_name: str,
    model_name: str,
    prompts: dict,
    correct: bool = False,
    split_correction: bool = False,
    version: str | None = None,
    base_url: str | None = None,
    update_token_usage: Callable | None = None,
) -> None:
    """Connect to Azure's GPT API and set up a conversation with the user.

    Extends GptConversation.

    Args:
    ----
        deployment_name (str): The name of the Azure deployment to use.

        model_name (str): The name of the model to use. This is distinct
            from the deployment name.

        prompts (dict): A dictionary of prompts to use for the conversation.

        correct (bool): Whether to correct the model output.

        split_correction (bool): Whether to correct the model output by
            splitting the output into sentences and correcting each
            sentence individually.

        version (str): The version of the Azure API to use.

        base_url (str): The base URL of the Azure API to use.

        update_token_usage (Callable): A function to update the token usage
            statistics.

    """
    super().__init__(
        model_name=model_name,
        prompts=prompts,
        correct=correct,
        split_correction=split_correction,
        update_token_usage=update_token_usage,
    )

    self.version = version
    self.base_url = base_url
    self.deployment_name = deployment_name

set_api_key(api_key, user=None)

Set the API key for the Azure API.

If the key is valid, initialise the conversational agent. No user stats on Azure.


api_key (str): The API key for the Azure API.

user (str, optional): The user for usage statistics.

bool: True if the API key is valid, False otherwise.
Source code in biochatter/llm_connect/azure.py
def set_api_key(self, api_key: str, user: str | None = None) -> bool:
    """Set the API key for the Azure API.

    If the key is valid, initialise the conversational agent. No user stats
    on Azure.

    Args:
    ----
        api_key (str): The API key for the Azure API.

        user (str, optional): The user for usage statistics.

    Returns:
    -------
        bool: True if the API key is valid, False otherwise.

    """
    try:
        self.chat = AzureChatOpenAI(
            deployment_name=self.deployment_name,
            model_name=self.model_name,
            openai_api_version=self.version,
            azure_endpoint=self.base_url,
            openai_api_key=api_key,
            temperature=0,
        )
        self.ca_chat = AzureChatOpenAI(
            deployment_name=self.deployment_name,
            model_name=self.model_name,
            openai_api_version=self.version,
            azure_endpoint=self.base_url,
            openai_api_key=api_key,
            temperature=0,
        )

        self.chat.generate([[HumanMessage(content="Hello")]])
        self.user = user if user is not None else "Azure Community"

        return True

    except openai._exceptions.AuthenticationError:
        self._chat = None
        self._ca_chat = None
        return False

BloomConversation

Bases: Conversation

Conversation class for the Bloom model.

Source code in biochatter/llm_connect/misc.py
class BloomConversation(Conversation):
    """Conversation class for the Bloom model."""

    def __init__(
        self,
        model_name: str,
        prompts: dict,
        split_correction: bool,
    ) -> None:
        """Initialise the BloomConversation class.

        DEPRECATED: Superceded by XinferenceConversation.
        """
        super().__init__(
            model_name=model_name,
            prompts=prompts,
            split_correction=split_correction,
        )

        self.messages = []

    def set_api_key(self, api_key: str, user: str | None = None) -> bool:
        """Set the API key for the HuggingFace API.

        If the key is valid, initialise the conversational agent.

        Args:
        ----
            api_key (str): The API key for the HuggingFace API.

            user (str): The user for usage statistics.

        Returns:
        -------
            bool: True if the API key is valid, False otherwise.

        """
        self.chat = HuggingFaceHub(
            repo_id=self.model_name,
            model_kwargs={"temperature": 1.0},  # "regular sampling"
            # as per https://huggingface.co/docs/api-inference/detailed_parameters
            huggingfacehub_api_token=api_key,
        )

        try:
            self.chat.generate(["Hello, I am a biomedical researcher."])
            return True
        except ValueError:
            return False

    def _cast_messages(self, messages: list) -> str:
        """Render the different roles of the chat-based conversation."""
        cast = ""
        for m in messages:
            if isinstance(m, SystemMessage):
                cast += f"System: {m.content}\n"
            elif isinstance(m, HumanMessage):
                cast += f"Human: {m.content}\n"
            elif isinstance(m, AIMessage):
                cast += f"AI: {m.content}\n"
            else:
                error_msg = f"Unknown message type: {type(m)}"
                raise TypeError(error_msg)

        return cast

    def _primary_query(self) -> tuple:
        response = self.chat.generate([self._cast_messages(self.messages)])

        msg = response.generations[0][0].text
        token_usage_raw = {
            "prompt_tokens": 0,
            "completion_tokens": 0,
            "total_tokens": 0,
        }
        token_usage = self._extract_total_tokens(token_usage_raw)

        self.append_ai_message(msg)

        return msg, token_usage

    def _correct_response(self, msg: str) -> str:
        return "ok"

__init__(model_name, prompts, split_correction)

Initialise the BloomConversation class.

DEPRECATED: Superceded by XinferenceConversation.

Source code in biochatter/llm_connect/misc.py
def __init__(
    self,
    model_name: str,
    prompts: dict,
    split_correction: bool,
) -> None:
    """Initialise the BloomConversation class.

    DEPRECATED: Superceded by XinferenceConversation.
    """
    super().__init__(
        model_name=model_name,
        prompts=prompts,
        split_correction=split_correction,
    )

    self.messages = []

set_api_key(api_key, user=None)

Set the API key for the HuggingFace API.

If the key is valid, initialise the conversational agent.


api_key (str): The API key for the HuggingFace API.

user (str): The user for usage statistics.

bool: True if the API key is valid, False otherwise.
Source code in biochatter/llm_connect/misc.py
def set_api_key(self, api_key: str, user: str | None = None) -> bool:
    """Set the API key for the HuggingFace API.

    If the key is valid, initialise the conversational agent.

    Args:
    ----
        api_key (str): The API key for the HuggingFace API.

        user (str): The user for usage statistics.

    Returns:
    -------
        bool: True if the API key is valid, False otherwise.

    """
    self.chat = HuggingFaceHub(
        repo_id=self.model_name,
        model_kwargs={"temperature": 1.0},  # "regular sampling"
        # as per https://huggingface.co/docs/api-inference/detailed_parameters
        huggingfacehub_api_token=api_key,
    )

    try:
        self.chat.generate(["Hello, I am a biomedical researcher."])
        return True
    except ValueError:
        return False

Conversation

Bases: ABC

Use this class to set up a connection to an LLM API.

Can be used to set the user name and API key, append specific messages for system, user, and AI roles (if available), set up the general context as well as manual and tool-based data inputs, and finally to query the API with prompts made by the user.

The conversation class is expected to have a messages attribute to store the conversation, and a history attribute, which is a list of messages in a specific format for logging / printing.

Source code in biochatter/llm_connect/conversation.py
  88
  89
  90
  91
  92
  93
  94
  95
  96
  97
  98
  99
 100
 101
 102
 103
 104
 105
 106
 107
 108
 109
 110
 111
 112
 113
 114
 115
 116
 117
 118
 119
 120
 121
 122
 123
 124
 125
 126
 127
 128
 129
 130
 131
 132
 133
 134
 135
 136
 137
 138
 139
 140
 141
 142
 143
 144
 145
 146
 147
 148
 149
 150
 151
 152
 153
 154
 155
 156
 157
 158
 159
 160
 161
 162
 163
 164
 165
 166
 167
 168
 169
 170
 171
 172
 173
 174
 175
 176
 177
 178
 179
 180
 181
 182
 183
 184
 185
 186
 187
 188
 189
 190
 191
 192
 193
 194
 195
 196
 197
 198
 199
 200
 201
 202
 203
 204
 205
 206
 207
 208
 209
 210
 211
 212
 213
 214
 215
 216
 217
 218
 219
 220
 221
 222
 223
 224
 225
 226
 227
 228
 229
 230
 231
 232
 233
 234
 235
 236
 237
 238
 239
 240
 241
 242
 243
 244
 245
 246
 247
 248
 249
 250
 251
 252
 253
 254
 255
 256
 257
 258
 259
 260
 261
 262
 263
 264
 265
 266
 267
 268
 269
 270
 271
 272
 273
 274
 275
 276
 277
 278
 279
 280
 281
 282
 283
 284
 285
 286
 287
 288
 289
 290
 291
 292
 293
 294
 295
 296
 297
 298
 299
 300
 301
 302
 303
 304
 305
 306
 307
 308
 309
 310
 311
 312
 313
 314
 315
 316
 317
 318
 319
 320
 321
 322
 323
 324
 325
 326
 327
 328
 329
 330
 331
 332
 333
 334
 335
 336
 337
 338
 339
 340
 341
 342
 343
 344
 345
 346
 347
 348
 349
 350
 351
 352
 353
 354
 355
 356
 357
 358
 359
 360
 361
 362
 363
 364
 365
 366
 367
 368
 369
 370
 371
 372
 373
 374
 375
 376
 377
 378
 379
 380
 381
 382
 383
 384
 385
 386
 387
 388
 389
 390
 391
 392
 393
 394
 395
 396
 397
 398
 399
 400
 401
 402
 403
 404
 405
 406
 407
 408
 409
 410
 411
 412
 413
 414
 415
 416
 417
 418
 419
 420
 421
 422
 423
 424
 425
 426
 427
 428
 429
 430
 431
 432
 433
 434
 435
 436
 437
 438
 439
 440
 441
 442
 443
 444
 445
 446
 447
 448
 449
 450
 451
 452
 453
 454
 455
 456
 457
 458
 459
 460
 461
 462
 463
 464
 465
 466
 467
 468
 469
 470
 471
 472
 473
 474
 475
 476
 477
 478
 479
 480
 481
 482
 483
 484
 485
 486
 487
 488
 489
 490
 491
 492
 493
 494
 495
 496
 497
 498
 499
 500
 501
 502
 503
 504
 505
 506
 507
 508
 509
 510
 511
 512
 513
 514
 515
 516
 517
 518
 519
 520
 521
 522
 523
 524
 525
 526
 527
 528
 529
 530
 531
 532
 533
 534
 535
 536
 537
 538
 539
 540
 541
 542
 543
 544
 545
 546
 547
 548
 549
 550
 551
 552
 553
 554
 555
 556
 557
 558
 559
 560
 561
 562
 563
 564
 565
 566
 567
 568
 569
 570
 571
 572
 573
 574
 575
 576
 577
 578
 579
 580
 581
 582
 583
 584
 585
 586
 587
 588
 589
 590
 591
 592
 593
 594
 595
 596
 597
 598
 599
 600
 601
 602
 603
 604
 605
 606
 607
 608
 609
 610
 611
 612
 613
 614
 615
 616
 617
 618
 619
 620
 621
 622
 623
 624
 625
 626
 627
 628
 629
 630
 631
 632
 633
 634
 635
 636
 637
 638
 639
 640
 641
 642
 643
 644
 645
 646
 647
 648
 649
 650
 651
 652
 653
 654
 655
 656
 657
 658
 659
 660
 661
 662
 663
 664
 665
 666
 667
 668
 669
 670
 671
 672
 673
 674
 675
 676
 677
 678
 679
 680
 681
 682
 683
 684
 685
 686
 687
 688
 689
 690
 691
 692
 693
 694
 695
 696
 697
 698
 699
 700
 701
 702
 703
 704
 705
 706
 707
 708
 709
 710
 711
 712
 713
 714
 715
 716
 717
 718
 719
 720
 721
 722
 723
 724
 725
 726
 727
 728
 729
 730
 731
 732
 733
 734
 735
 736
 737
 738
 739
 740
 741
 742
 743
 744
 745
 746
 747
 748
 749
 750
 751
 752
 753
 754
 755
 756
 757
 758
 759
 760
 761
 762
 763
 764
 765
 766
 767
 768
 769
 770
 771
 772
 773
 774
 775
 776
 777
 778
 779
 780
 781
 782
 783
 784
 785
 786
 787
 788
 789
 790
 791
 792
 793
 794
 795
 796
 797
 798
 799
 800
 801
 802
 803
 804
 805
 806
 807
 808
 809
 810
 811
 812
 813
 814
 815
 816
 817
 818
 819
 820
 821
 822
 823
 824
 825
 826
 827
 828
 829
 830
 831
 832
 833
 834
 835
 836
 837
 838
 839
 840
 841
 842
 843
 844
 845
 846
 847
 848
 849
 850
 851
 852
 853
 854
 855
 856
 857
 858
 859
 860
 861
 862
 863
 864
 865
 866
 867
 868
 869
 870
 871
 872
 873
 874
 875
 876
 877
 878
 879
 880
 881
 882
 883
 884
 885
 886
 887
 888
 889
 890
 891
 892
 893
 894
 895
 896
 897
 898
 899
 900
 901
 902
 903
 904
 905
 906
 907
 908
 909
 910
 911
 912
 913
 914
 915
 916
 917
 918
 919
 920
 921
 922
 923
 924
 925
 926
 927
 928
 929
 930
 931
 932
 933
 934
 935
 936
 937
 938
 939
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
class Conversation(ABC):
    """Use this class to set up a connection to an LLM API.

    Can be used to set the user name and API key, append specific messages for
    system, user, and AI roles (if available), set up the general context as
    well as manual and tool-based data inputs, and finally to query the API
    with prompts made by the user.

    The conversation class is expected to have a `messages` attribute to store
    the conversation, and a `history` attribute, which is a list of messages in
    a specific format for logging / printing.

    """

    def __init__(
        self,
        model_name: str,
        prompts: dict,
        correct: bool = False,
        split_correction: bool = False,
        use_ragagent_selector: bool = False,
        tools: list[Callable] = None,
        tool_call_mode: Literal["auto", "text"] = "auto",
        mcp: bool = False,
        additional_tools_instructions: str = None,
        force_tool: bool = False,
    ) -> None:
        super().__init__()
        self.model_name = model_name
        self.prompts = prompts
        self.correct = correct
        self.split_correction = split_correction
        self.rag_agents: list[RagAgent] = []
        self.history = []
        self.messages = []
        self.ca_messages = []
        self.tool_calls = deque()
        self.current_statements = []
        self._use_ragagent_selector = use_ragagent_selector
        self._chat = None
        self._ca_chat = None
        self.tools = tools
        self.tool_call_mode = tool_call_mode
        self.tools_prompt = None
        self.mcp = mcp
        self.additional_tools_instructions = additional_tools_instructions if additional_tools_instructions else ""
        self.force_tool = force_tool

    @property
    def chat(self):
        """Access the chat attribute with error handling."""
        if self._chat is None:
            msg = "Chat attribute not initialized. Did you call set_api_key()?"
            logger.error(msg)
            raise AttributeError(msg)
        return self._chat

    @chat.setter
    def chat(self, value):
        """Set the chat attribute."""
        self._chat = value

    @property
    def ca_chat(self):
        """Access the correcting agent chat attribute with error handling."""
        if self._ca_chat is None:
            msg = "Correcting agent chat attribute not initialized. Did you call set_api_key()?"
            logger.error(msg)
            raise AttributeError(msg)
        return self._ca_chat

    @ca_chat.setter
    def ca_chat(self, value):
        """Set the correcting agent chat attribute."""
        self._ca_chat = value

    @property
    def use_ragagent_selector(self) -> bool:
        """Whether to use the ragagent selector."""
        return self._use_ragagent_selector

    @use_ragagent_selector.setter
    def use_ragagent_selector(self, val: bool) -> None:
        """Set the use_ragagent_selector attribute."""
        self._use_ragagent_selector = val

    def set_user_name(self, user_name: str) -> None:
        """Set the user name."""
        self.user_name = user_name

    def set_rag_agent(self, agent: RagAgent) -> None:
        """Update or insert rag_agent.

        If the rag_agent with the same mode already exists, it will be updated.
        Otherwise, the new rag_agent will be inserted.
        """
        i, _ = self.find_rag_agent(agent.mode)
        if i < 0:
            # insert
            self.rag_agents.append(agent)
        else:
            # update
            self.rag_agents[i] = agent

    def find_rag_agent(self, mode: str) -> tuple[int, RagAgent]:
        """Find the rag_agent with the given mode."""
        for i, val in enumerate(self.rag_agents):
            if val.mode == mode:
                return i, val
        return -1, None

    def _extract_total_tokens(self, token_usage: dict | int | None) -> int | None:
        """Extract total tokens from various token usage formats.

        This method standardizes token counting across different providers:
        - OpenAI/Azure: {"prompt_tokens": X, "completion_tokens": Y, "total_tokens": Z}
        - Anthropic: {"input_tokens": X, "output_tokens": Y} -> calculate total
        - Gemini: {"total_tokens": Z} -> extract total
        - Ollama: integer (eval_count) -> return as is
        - LiteLLM: {"input_tokens": X, "output_tokens": Y, "total_tokens": Z}
        - Others: try to extract or calculate total

        Args:
        ----
            token_usage: Token usage in various formats (dict, int, or None)

        Returns:
        -------
            int | None: Total token count, or None if not available

        """
        if token_usage is None:
            return None

        # Handle integer token counts (Ollama, some others)
        if isinstance(token_usage, int):
            return token_usage

        # Handle dictionary token counts
        if isinstance(token_usage, dict):
            # First try to get total_tokens directly
            if "total_tokens" in token_usage:
                return token_usage["total_tokens"]

            # Calculate from input/output tokens (Anthropic style)
            if "input_tokens" in token_usage and "output_tokens" in token_usage:
                return token_usage["input_tokens"] + token_usage["output_tokens"]

            # Calculate from prompt/completion tokens (OpenAI style fallback)
            if "prompt_tokens" in token_usage and "completion_tokens" in token_usage:
                return token_usage["prompt_tokens"] + token_usage["completion_tokens"]

            # If only one type of token count is available, use it
            if "input_tokens" in token_usage:
                return token_usage["input_tokens"]
            if "output_tokens" in token_usage:
                return token_usage["output_tokens"]
            if "prompt_tokens" in token_usage:
                return token_usage["prompt_tokens"]
            if "completion_tokens" in token_usage:
                return token_usage["completion_tokens"]

        # If we can't extract meaningful token count, return None
        return None

    def _extract_input_tokens(self, token_usage: dict | int | None) -> int | None:
        """Extract input tokens from various token usage formats.

        This method standardizes input token counting across different providers:
        - OpenAI/Azure: {"prompt_tokens": X, "completion_tokens": Y, "total_tokens": Z}
        - Anthropic: {"input_tokens": X, "output_tokens": Y}
        - Gemini: {"prompt_tokens": X, "candidates_tokens": Y, "total_tokens": Z}
        - LiteLLM: {"input_tokens": X, "output_tokens": Y, "total_tokens": Z}
        - Others: try to extract input/prompt tokens

        Args:
        ----
            token_usage: Token usage in various formats (dict, int, or None)

        Returns:
        -------
            int | None: Input token count, or None if not available

        """
        if token_usage is None:
            return None

        # Handle integer token counts (cannot distinguish input vs output)
        if isinstance(token_usage, int):
            return None

        # Handle dictionary token counts
        if isinstance(token_usage, dict):
            # First try to get input_tokens (Anthropic, LiteLLM style)
            if "input_tokens" in token_usage:
                return token_usage["input_tokens"]

            # Try prompt_tokens (OpenAI style)
            if "prompt_tokens" in token_usage:
                return token_usage["prompt_tokens"]

        # If we can't extract meaningful input token count, return None
        return None

    def _extract_output_tokens(self, token_usage: dict | int | None) -> int | None:
        """Extract output tokens from various token usage formats.

        This method standardizes output token counting across different providers:
        - OpenAI/Azure: {"prompt_tokens": X, "completion_tokens": Y, "total_tokens": Z}
        - Anthropic: {"input_tokens": X, "output_tokens": Y}
        - Gemini: {"prompt_tokens": X, "candidates_tokens": Y, "total_tokens": Z}
        - LiteLLM: {"input_tokens": X, "output_tokens": Y, "total_tokens": Z}
        - Others: try to extract output/completion tokens

        Args:
        ----
            token_usage: Token usage in various formats (dict, int, or None)

        Returns:
        -------
            int | None: Output token count, or None if not available

        """
        if token_usage is None:
            return None

        # Handle integer token counts (cannot distinguish input vs output)
        if isinstance(token_usage, int):
            return None

        # Handle dictionary token counts
        if isinstance(token_usage, dict):
            # First try to get output_tokens (Anthropic, LiteLLM style)
            if "output_tokens" in token_usage:
                return token_usage["output_tokens"]

            # Try completion_tokens (OpenAI style)
            if "completion_tokens" in token_usage:
                return token_usage["completion_tokens"]

            # Try candidates_tokens (Gemini style)
            if "candidates_tokens" in token_usage:
                return token_usage["candidates_tokens"]

        # If we can't extract meaningful output token count, return None
        return None

    def compute_cumulative_token_usage(self) -> dict:
        """Compute the token usage by looping over the messages.

        Extracts token usage information from each message's usage_metadata and
        computes running cumulative totals throughout the conversation.
        Handles various token usage formats from different LLM providers.

        Returns
        -------
            dict: Token usage information with lists of running totals:
                - "total_tokens": list[int] - running total at each message
                - "input_tokens": list[int] - running input total at each message
                - "output_tokens": list[int] - running output total at each message

        """
        # Initialize data structures
        individual_usage = {
            "total_tokens": [],
            "input_tokens": [],
            "output_tokens": [],
        }

        # Extract individual token counts for each AI message
        for message in self.messages:
            if isinstance(message, AIMessage):
                usage_metadata = getattr(message, "usage_metadata", None)
                individual_usage["total_tokens"].append(self._extract_total_tokens(usage_metadata))
                individual_usage["input_tokens"].append(self._extract_input_tokens(usage_metadata))
                individual_usage["output_tokens"].append(self._extract_output_tokens(usage_metadata))

        # Compute running cumulative totals for each message
        per_message_cumulative = {
            "total_tokens": [],
            "input_tokens": [],
            "output_tokens": [],
        }

        for token_type in ["total_tokens", "input_tokens", "output_tokens"]:
            running_total = 0
            for count in individual_usage[token_type]:
                if count is not None:
                    running_total += count
                per_message_cumulative[token_type].append(running_total)

        return per_message_cumulative

    @abstractmethod
    def set_api_key(self, api_key: str, user: str | None = None) -> None:
        """Set the API key."""

    def get_prompts(self) -> dict:
        """Get the prompts."""
        return self.prompts

    def set_prompts(self, prompts: dict) -> None:
        """Set the prompts."""
        self.prompts = prompts

    def _tool_formatter(self, tools: list[Callable], mcp: bool = False) -> str:
        """Format the tools. Only for model not supporting tool calling."""
        tools_description = ""

        for idx, tool in enumerate(tools):
            tools_description += f"<tool_{idx}>\n"
            tools_description += f"Tool name: {tool.name}\n"
            tools_description += f"Tool description: {tool.description}\n"
            if mcp:
                tools_description += f"Tool call schema:\n {tool.tool_call_schema}\n"
            else:
                tools_description += f"Tool call schema:\n {tool.args}\n"
            tools_description += f"</tool_{idx}>\n"
        return tools_description

    def _create_tool_prompt(
        self, tools: list[Callable], additional_tools_instructions: str = None, mcp: bool = False
    ) -> str:
        """Create the tool prompt. Only for model not supporting tool calling."""
        prompt_template = ChatPromptTemplate.from_template(TOOL_USAGE_PROMPT)
        tools_description = self._tool_formatter(tools, mcp=mcp)
        new_message = prompt_template.invoke(
            {
                "user_question": self.messages[-1].content,
                "tools": tools_description,
                "additional_tools_instructions": additional_tools_instructions if additional_tools_instructions else "",
            }
        )
        return new_message.messages[0]

    def bind_tools(self, tools: list[Callable]) -> None:
        """Bind tools to the chat."""
        # Check if the model supports tool calling
        # (exploit the enum class in available_models.py)
        if self.model_name in TOOL_CALLING_MODELS and self.ca_chat:
            self.chat = self.chat.bind_tools(tools)
            self.ca_chat = self.ca_chat.bind_tools(tools)

        elif self.model_name in TOOL_CALLING_MODELS:
            self.chat = self.chat.bind_tools(tools)

        # elif self.model_name not in TOOL_CALLING_MODELS:
        #    self.tools_prompt = self._create_tool_prompt(tools, additional_instructions)

        # If not, fail gracefully
        # raise ValueError(f"Model {self.model_name} does not support tool calling.")

    def append_ai_message(self, message: str | AIMessage) -> None:
        """Add a message from the AI to the conversation.

        Args:
        ----
            message (str): The message from the AI.

        """
        if isinstance(message, AIMessage):
            self.messages.append(message)
        elif isinstance(message, str):
            self.messages.append(
                AIMessage(
                    content=message,
                ),
            )
        else:
            raise ValueError(f"Invalid message type: {type(message)}")

    def append_system_message(self, message: str) -> None:
        """Add a system message to the conversation.

        Args:
        ----
            message (str): The system message.

        """
        self.messages.append(
            SystemMessage(
                content=message,
            ),
        )

    def append_ca_message(self, message: str) -> None:
        """Add a message to the correcting agent conversation.

        Args:
        ----
            message (str): The message to the correcting agent.

        """
        self.ca_messages.append(
            SystemMessage(
                content=message,
            ),
        )

    def append_user_message(self, message: str) -> None:
        """Add a message from the user to the conversation.

        Args:
        ----
            message (str): The message from the user.

        """
        self.messages.append(
            HumanMessage(
                content=message,
            ),
        )

    def append_image_message(
        self,
        message: str,
        image_url: str,
        local: bool = False,
    ) -> None:
        """Add a user message with an image to the conversation.

        Also checks, in addition to the `local` flag, if the image URL is a
        local file path. If it is local, the image will be encoded as a base64
        string to be passed to the LLM.

        Args:
        ----
            message (str): The message from the user.
            image_url (str): The URL of the image.
            local (bool): Whether the image is local or not. If local, it will
                be encoded as a base64 string to be passed to the LLM.

        """
        parsed_url = urllib.parse.urlparse(image_url)
        if local or not parsed_url.netloc:
            image_url = f"data:image/jpeg;base64,{encode_image(image_url)}"
        else:
            image_url = f"data:image/jpeg;base64,{encode_image_from_url(image_url)}"

        self.messages.append(
            HumanMessage(
                content=[
                    {"type": "text", "text": message},
                    {"type": "image_url", "image_url": {"url": image_url}},
                ],
            ),
        )

    def setup(self, context: str) -> None:
        """Set up the conversation with general prompts and a context."""
        for msg in self.prompts["primary_model_prompts"]:
            if msg:
                self.append_system_message(msg)

        for msg in self.prompts["correcting_agent_prompts"]:
            if msg:
                self.append_ca_message(msg)

        self.context = context
        msg = f"The topic of the research is {context}."
        self.append_system_message(msg)

    def setup_data_input_manual(self, data_input: str) -> None:
        """Set up the data input manually."""
        self.data_input = data_input
        msg = f"The user has given information on the data input: {data_input}."
        self.append_system_message(msg)

    def setup_data_input_tool(self, df, input_file_name: str) -> None:
        """Set up the data input tool."""
        self.data_input_tool = df

        for tool_name in self.prompts["tool_prompts"]:
            if tool_name in input_file_name:
                msg = self.prompts["tool_prompts"][tool_name].format(df=df)
                self.append_system_message(msg)

    def query(
        self,
        text: str,
        image_url: str | None = None,
        structured_model: BaseModel | None = None,
        wrap_structured_output: bool | None = None,
        tools: list[Callable] | None = None,
        explain_tool_result: bool | None = None,
        additional_tools_instructions: str | None = None,
        general_instructions_tool_interpretation: str | None = None,
        additional_instructions_tool_interpretation: str | None = None,
        mcp: bool | None = None,
        return_tool_calls_as_ai_message: bool | None = None,
        track_tool_calls: bool | None = None,
        **kwargs,
    ) -> tuple[str, dict | None, str | None]:
        """Query the LLM API using the user's query.

        Appends the most recent query to the conversation, optionally injects
        context from the RAG agent, and runs the primary query method of the
        child class.

        Args:
        ----
            text (str): The user query.

            image_url (str): The URL of an image to include in the conversation.
                Optional and only supported for models with vision capabilities.

            structured_model (BaseModel): The structured output model to use for the query.

            wrap_structured_output (bool): Whether to wrap the structured output in JSON quotes.

            tools (list[Callable]): The tools to use for the query.

            explain_tool_result (bool): Whether to explain the tool result.

            additional_tools_instructions (str): The additional instructions for the query.
                Mainly used for tools that do not support tool calling.

            general_instructions_tool_interpretation (str): The general
                instructions for the tool interpretation.
                Overrides the default prompt in `GENERAL_TOOL_RESULT_INTERPRETATION_PROMPT`.

            additional_instructions_tool_interpretation (str): The additional
                instructions for the tool interpretation.
                Overrides the default prompt in `ADDITIONAL_TOOL_RESULT_INTERPRETATION_PROMPT`.

            mcp (bool): If you want to use MCP mode, this should be set to True.

            return_tool_calls_as_ai_message (bool): If you want to return the tool calls as an AI message, this should be set to True.

            track_tool_calls (bool): If you want to track the tool calls, this should be set to True.

            **kwargs: Additional keyword arguments.

        Returns:
        -------
            tuple: A tuple containing the response from the API, the token usage
                information, and the correction if necessary/desired.

        """
        if mcp:
            self.mcp = True

        # save the last human prompt that may be used for answer enhancement
        self.last_human_prompt = text

        # if additional_tools_instructions are provided, save them
        if additional_tools_instructions:
            self.additional_tools_instructions = additional_tools_instructions

        # override the default prompts if other provided
        self.general_instructions_tool_interpretation = (
            general_instructions_tool_interpretation
            if general_instructions_tool_interpretation
            else GENERAL_TOOL_RESULT_INTERPRETATION_PROMPT
        )
        self.additional_instructions_tool_interpretation = (
            additional_instructions_tool_interpretation
            if additional_instructions_tool_interpretation
            else ADDITIONAL_TOOL_RESULT_INTERPRETATION_PROMPT
        )
        if not image_url:
            self.append_user_message(text)
        else:
            self.append_image_message(text, image_url)

        self._inject_context(text)

        # tools passed at this step are used only for this message
        msg, token_usage = self._primary_query(
            tools=tools,
            explain_tool_result=explain_tool_result,
            return_tool_calls_as_ai_message=return_tool_calls_as_ai_message,
            structured_model=structured_model,
            wrap_structured_output=wrap_structured_output,
            track_tool_calls=track_tool_calls,
        )

        # case of structured output
        if (token_usage == -1) and structured_model:
            return (msg, 0, None)

        if not token_usage:
            # indicates error
            return (msg, None, None)

        if not self.correct:
            return (msg, token_usage, None)

        cor_msg = "Correcting (using single sentences) ..." if self.split_correction else "Correcting ..."

        if st:
            with st.spinner(cor_msg):
                corrections = self._correct_query(text)
        else:
            corrections = self._correct_query(text)

        if not corrections:
            return (msg, token_usage, None)

        correction = "\n".join(corrections)
        return (msg, token_usage, correction)

    def _correct_query(self, msg: str) -> list[str]:
        corrections = []
        if self.split_correction:
            nltk.download("punkt")
            tokenizer = nltk.data.load("tokenizers/punkt/english.pickle")
            sentences = tokenizer.tokenize(msg)
            for sentence in sentences:
                correction = self._correct_response(sentence)

                if str(correction).lower() not in ["ok", "ok."]:
                    corrections.append(correction)
        else:
            correction = self._correct_response(msg)

            if str(correction).lower() not in ["ok", "ok."]:
                corrections.append(correction)

        return corrections

    @abstractmethod
    def _primary_query(self, **kwargs) -> tuple[str, dict | None]:
        """Run the primary query.

        Args:
        ----
            **kwargs: Keyword arguments that may include:
                - text: The user query.
                - tools: List of tools for tool-calling models
                - explain_tool_result: Whether to explain tool results
                - return_tool_calls_as_ai_message: Whether to return tool calls as AI message
                - structured_model: Structured output model
                - wrap_structured_output: Whether to wrap structured output
                - track_tool_calls: Whether to track tool calls
                - Other model-specific parameters

        Returns:
        -------
            tuple: A tuple containing the response message and token usage information.

        """

    @abstractmethod
    def _correct_response(self, msg: str) -> str:
        """Correct the response."""

    def _process_manual_tool_call(
        self,
        tool_call: list[dict],
        available_tools: list[Callable],
        explain_tool_result: bool = False,
    ) -> str:
        """Process manual tool calls from the model response.

        This method handles the processing of tool calls for models that don't natively
        support tool calling. It takes the parsed JSON response and executes the
        appropriate tool.

        Args:
        ----
            tool_call (list[dict]): The parsed tool call information from the model response.
            available_tools (list[Callable]): The tools available for execution.
            explain_tool_result (bool): Whether to explain the tool result.

        Returns:
        -------
            str: The processed message containing the tool name, arguments, and result.

        """
        tool_name = tool_call["tool_name"]
        tool_func = next((t for t in available_tools if t.name == tool_name), None)

        # Remove the tool name from the tool call in order to invoke the tool
        # This is beacause tool_name is not a valid argument for the tool
        del tool_call["tool_name"]

        # Execute the tool based on whether we're in async context or not
        if self.mcp:
            loop = asyncio.get_running_loop()
            tool_result = loop.run_until_complete(tool_func.ainvoke(tool_call))
        else:
            tool_result = tool_func.invoke(tool_call)

        msg = f"Tool: {tool_name}\nArguments: {tool_call}\nTool result: {tool_result}"

        if explain_tool_result:
            tool_result_interpretation = self.chat.invoke(
                TOOL_RESULT_INTERPRETATION_PROMPT.format(
                    original_question=self.last_human_prompt,
                    tool_result=tool_result,
                    general_instructions=self.general_instructions_tool_interpretation,
                    additional_instructions=self.additional_instructions_tool_interpretation,
                )
            )
            msg += f"\nTool result interpretation: {tool_result_interpretation.content}"

        self.append_ai_message(msg)

        return msg

    def _process_tool_calls(
        self,
        tool_calls: list[dict],
        available_tools: list[Callable],
        response_content: str,
        explain_tool_result: bool = False,
        return_tool_calls_as_ai_message: bool = False,
        track_tool_calls: bool = False,
    ) -> str:
        """Process tool calls from the model response.

        This method handles the processing of tool calls returned by the model.
        It can either automatically execute the tools and return their results,
        or format the tool calls as text.

        Args:
        ----
            tool_calls: The tool calls from the model response.
            response_content: The text content of the response (used as fallback).
            available_tools: The tools available in the chat.
            explain_tool_result (bool): Whether to explain the tool result.
            return_tool_calls_as_ai_message (bool): If you want to return the tool calls as an AI message, this should be set to True.
            track_tool_calls (bool): If you want to track the tool calls, this should be set to True.

        Returns:
        -------
            str: The processed message, either tool results or formatted tool calls.

        """
        if not tool_calls:
            return response_content

        msg = ""

        if self.tool_call_mode == "auto":
            # Collect tool results for collective explanation when multiple tools are called
            tool_results_for_explanation = []

            for idx, tool_call in enumerate(tool_calls):
                # Extract tool name and arguments
                tool_name = tool_call["name"]
                tool_args = tool_call["args"]
                tool_call_id = tool_call["id"]

                # Find the matching tool function
                tool_func = next((t for t in available_tools if t.name == tool_name), None)

                if tool_func:
                    # Execute the tool
                    try:
                        if self.mcp:
                            loop = asyncio.get_running_loop()
                            tool_result = loop.run_until_complete(tool_func.ainvoke(tool_args))
                        else:
                            tool_result = tool_func.invoke(tool_args)
                        # Add the tool result to the conversation
                        if return_tool_calls_as_ai_message:
                            self.append_ai_message(f"Tool call ({tool_name}) \nResult: {tool_result!s}")
                        else:
                            self.messages.append(
                                ToolMessage(content=str(tool_result), name=tool_name, tool_call_id=tool_call_id)
                            )

                        if track_tool_calls:
                            self.tool_calls.append(
                                {"name": tool_name, "args": tool_args, "id": tool_call_id, "result": tool_result}
                            )

                        if idx > 0:
                            msg += "\n"
                        msg += f"Tool call ({tool_name}) result: {tool_result!s}"

                        # Collect tool results for explanation if needed
                        if explain_tool_result:
                            tool_results_for_explanation.append(
                                {"name": tool_name, "args": tool_args, "result": tool_result}
                            )

                    except Exception as e:
                        # Handle tool execution errors
                        error_message = f"Error executing tool {tool_name}: {e!s}"
                        self.messages.append(
                            ToolMessage(content=error_message, name=tool_name, tool_call_id=tool_call_id)
                        )

                        # Track failed tool calls
                        if track_tool_calls:
                            self.tool_calls.append(
                                {"name": tool_name, "args": tool_args, "id": tool_call_id, "error": str(e)}
                            )

                        if idx > 0:
                            msg += "\n"
                        msg += error_message
                # Handle missing/unknown tool
                elif track_tool_calls:
                    self.tool_calls.append(
                        {"name": tool_name, "args": tool_args, "id": tool_call_id, "error": "Tool not found"}
                    )

            # Handle tool result explanation
            if explain_tool_result and tool_results_for_explanation:
                if len(tool_results_for_explanation) > 1:
                    # Multiple tools: explain all results together
                    combined_tool_results = "\n\n".join(
                        [
                            f"Tool: {tr['name']}\nArguments: {tr['args']}\nResult: {tr['result']}"
                            for tr in tool_results_for_explanation
                        ]
                    )

                    tool_result_interpretation = self.chat.invoke(
                        TOOL_RESULT_INTERPRETATION_PROMPT.format(
                            original_question=self.last_human_prompt,
                            tool_result=combined_tool_results,
                            general_instructions=self.general_instructions_tool_interpretation,
                            additional_instructions=self.additional_instructions_tool_interpretation,
                        )
                    )
                    self.messages.append(tool_result_interpretation)
                    msg += f"\nTool results interpretation: {tool_result_interpretation.content}"
                else:
                    # Single tool: explain individual result (maintain current behavior)
                    tool_result_data = tool_results_for_explanation[0]
                    tool_result_interpretation = self.chat.invoke(
                        TOOL_RESULT_INTERPRETATION_PROMPT.format(
                            original_question=self.last_human_prompt,
                            tool_result=tool_result_data["result"],
                            general_instructions=self.general_instructions_tool_interpretation,
                            additional_instructions=self.additional_instructions_tool_interpretation,
                        )
                    )
                    self.messages.append(tool_result_interpretation)
                    msg += f"\nTool result interpretation: {tool_result_interpretation.content}"

            return msg

        if self.tool_call_mode == "text":
            # Join all tool calls in a text format
            tool_calls_text = []
            for tool_call in tool_calls:
                tool_name = tool_call["name"]
                tool_args = tool_call["args"]
                tool_call_id = tool_call["id"]
                tool_calls_text.append(f"Tool: {tool_name} - Arguments: {tool_args} - Tool call id: {tool_call_id}")

            # Join with line breaks and set as the message
            msg = "\n".join(tool_calls_text)

            # Append the formatted tool calls as an AI message
            self.append_ai_message(msg)
            return msg

        # Invalid tool call mode, log warning and return original content
        logger.warning(f"Invalid tool call mode: {self.tool_call_mode}. Using original response content.")
        return response_content

    def _inject_context_by_ragagent_selector(self, text: str) -> list[str]:
        """Inject the context generated by RagAgentSelector.

        The RagAgentSelector will choose the appropriate rag agent to generate
        context according to user's question.

        Args:
        ----
            text (str): The user query to be used for choosing rag agent

        """
        rag_agents: list[RagAgent] = [agent for agent in self.rag_agents if agent.use_prompt]
        decider_agent = RagAgentSelector(
            rag_agents=rag_agents,
            conversation_factory=lambda: self,
        )
        result = decider_agent.execute(text)
        if result.tool_result is not None and len(result.tool_result) > 0:
            return result.tool_result
        # find rag agent selected
        rag_agent = next(
            [agent for agent in rag_agents if agent.mode == result.answer],
            None,
        )
        if rag_agent is None:
            return None
        return rag_agent.generate_responses(text)

    def _inject_context(self, text: str) -> None:
        """Inject the context received from the RAG agent into the prompt.

        The RAG agent will find the most similar n text fragments and add them
        to the message history object for usage in the next prompt. Uses the
        document summarisation prompt set to inject the context. The ultimate
        prompt should include the placeholder for the statements, `{statements}`
        (used for formatting the string).

        Args:
        ----
            text (str): The user query to be used for similarity search.

        """
        sim_msg = "Performing similarity search to inject fragments ..."

        if st:
            with st.spinner(sim_msg):
                statements = []
                if self.use_ragagent_selector:
                    statements = self._inject_context_by_ragagent_selector(text)
                else:
                    for agent in self.rag_agents:
                        try:
                            docs = agent.generate_responses(text)
                            statements = statements + [doc[0] for doc in docs]
                        except ValueError as e:
                            logger.warning(e)

        else:
            statements = []
            if self.use_ragagent_selector:
                statements = self._inject_context_by_ragagent_selector(text)
            else:
                for agent in self.rag_agents:
                    try:
                        docs = agent.generate_responses(text)
                        statements = statements + [doc[0] for doc in docs]
                    except ValueError as e:
                        logger.warning(e)

        if statements and len(statements) > 0:
            prompts = self.prompts["rag_agent_prompts"]
            self.current_statements = statements
            for i, prompt in enumerate(prompts):
                # if last prompt, format the statements into the prompt
                if i == len(prompts) - 1:
                    self.append_system_message(
                        prompt.format(statements=statements),
                    )
                else:
                    self.append_system_message(prompt)

    def get_last_injected_context(self) -> list[dict]:
        """Get a formatted list of the last context.

        Get the last context injected into the conversation. Contains one
        dictionary for each RAG mode.

        Returns
        -------
            List[dict]: A list of dictionaries containing the mode and context
            for each RAG agent.

        """
        return [{"mode": agent.mode, "context": agent.last_response} for agent in self.rag_agents]

    def get_msg_json(self) -> str:
        """Return a JSON representation of the conversation.

        Returns a list of dicts of the messages in the conversation in JSON
        format. The keys of the dicts are the roles, the values are the
        messages.

        Returns
        -------
            str: A JSON representation of the messages in the conversation.

        """
        d = []
        for msg in self.messages:
            if isinstance(msg, SystemMessage):
                role = "system"
            elif isinstance(msg, HumanMessage):
                role = "user"
            elif isinstance(msg, AIMessage):
                role = "ai"
            else:
                error_msg = f"Unknown message type: {type(msg)}"
                raise TypeError(error_msg)

            d.append({role: msg.content})

        return json.dumps(d)

    def reset(self) -> None:
        """Reset the conversation to the initial state."""
        self.history = []
        self.messages = []
        self.ca_messages = []
        self.current_statements = []
        self.tool_calls.clear()

ca_chat property writable

Access the correcting agent chat attribute with error handling.

chat property writable

Access the chat attribute with error handling.

use_ragagent_selector property writable

Whether to use the ragagent selector.

append_ai_message(message)

Add a message from the AI to the conversation.


message (str): The message from the AI.
Source code in biochatter/llm_connect/conversation.py
def append_ai_message(self, message: str | AIMessage) -> None:
    """Add a message from the AI to the conversation.

    Args:
    ----
        message (str): The message from the AI.

    """
    if isinstance(message, AIMessage):
        self.messages.append(message)
    elif isinstance(message, str):
        self.messages.append(
            AIMessage(
                content=message,
            ),
        )
    else:
        raise ValueError(f"Invalid message type: {type(message)}")

append_ca_message(message)

Add a message to the correcting agent conversation.


message (str): The message to the correcting agent.
Source code in biochatter/llm_connect/conversation.py
def append_ca_message(self, message: str) -> None:
    """Add a message to the correcting agent conversation.

    Args:
    ----
        message (str): The message to the correcting agent.

    """
    self.ca_messages.append(
        SystemMessage(
            content=message,
        ),
    )

append_image_message(message, image_url, local=False)

Add a user message with an image to the conversation.

Also checks, in addition to the local flag, if the image URL is a local file path. If it is local, the image will be encoded as a base64 string to be passed to the LLM.


message (str): The message from the user.
image_url (str): The URL of the image.
local (bool): Whether the image is local or not. If local, it will
    be encoded as a base64 string to be passed to the LLM.
Source code in biochatter/llm_connect/conversation.py
def append_image_message(
    self,
    message: str,
    image_url: str,
    local: bool = False,
) -> None:
    """Add a user message with an image to the conversation.

    Also checks, in addition to the `local` flag, if the image URL is a
    local file path. If it is local, the image will be encoded as a base64
    string to be passed to the LLM.

    Args:
    ----
        message (str): The message from the user.
        image_url (str): The URL of the image.
        local (bool): Whether the image is local or not. If local, it will
            be encoded as a base64 string to be passed to the LLM.

    """
    parsed_url = urllib.parse.urlparse(image_url)
    if local or not parsed_url.netloc:
        image_url = f"data:image/jpeg;base64,{encode_image(image_url)}"
    else:
        image_url = f"data:image/jpeg;base64,{encode_image_from_url(image_url)}"

    self.messages.append(
        HumanMessage(
            content=[
                {"type": "text", "text": message},
                {"type": "image_url", "image_url": {"url": image_url}},
            ],
        ),
    )

append_system_message(message)

Add a system message to the conversation.


message (str): The system message.
Source code in biochatter/llm_connect/conversation.py
def append_system_message(self, message: str) -> None:
    """Add a system message to the conversation.

    Args:
    ----
        message (str): The system message.

    """
    self.messages.append(
        SystemMessage(
            content=message,
        ),
    )

append_user_message(message)

Add a message from the user to the conversation.


message (str): The message from the user.
Source code in biochatter/llm_connect/conversation.py
def append_user_message(self, message: str) -> None:
    """Add a message from the user to the conversation.

    Args:
    ----
        message (str): The message from the user.

    """
    self.messages.append(
        HumanMessage(
            content=message,
        ),
    )

bind_tools(tools)

Bind tools to the chat.

Source code in biochatter/llm_connect/conversation.py
def bind_tools(self, tools: list[Callable]) -> None:
    """Bind tools to the chat."""
    # Check if the model supports tool calling
    # (exploit the enum class in available_models.py)
    if self.model_name in TOOL_CALLING_MODELS and self.ca_chat:
        self.chat = self.chat.bind_tools(tools)
        self.ca_chat = self.ca_chat.bind_tools(tools)

    elif self.model_name in TOOL_CALLING_MODELS:
        self.chat = self.chat.bind_tools(tools)

compute_cumulative_token_usage()

Compute the token usage by looping over the messages.

Extracts token usage information from each message's usage_metadata and computes running cumulative totals throughout the conversation. Handles various token usage formats from different LLM providers.

Returns
dict: Token usage information with lists of running totals:
    - "total_tokens": list[int] - running total at each message
    - "input_tokens": list[int] - running input total at each message
    - "output_tokens": list[int] - running output total at each message
Source code in biochatter/llm_connect/conversation.py
def compute_cumulative_token_usage(self) -> dict:
    """Compute the token usage by looping over the messages.

    Extracts token usage information from each message's usage_metadata and
    computes running cumulative totals throughout the conversation.
    Handles various token usage formats from different LLM providers.

    Returns
    -------
        dict: Token usage information with lists of running totals:
            - "total_tokens": list[int] - running total at each message
            - "input_tokens": list[int] - running input total at each message
            - "output_tokens": list[int] - running output total at each message

    """
    # Initialize data structures
    individual_usage = {
        "total_tokens": [],
        "input_tokens": [],
        "output_tokens": [],
    }

    # Extract individual token counts for each AI message
    for message in self.messages:
        if isinstance(message, AIMessage):
            usage_metadata = getattr(message, "usage_metadata", None)
            individual_usage["total_tokens"].append(self._extract_total_tokens(usage_metadata))
            individual_usage["input_tokens"].append(self._extract_input_tokens(usage_metadata))
            individual_usage["output_tokens"].append(self._extract_output_tokens(usage_metadata))

    # Compute running cumulative totals for each message
    per_message_cumulative = {
        "total_tokens": [],
        "input_tokens": [],
        "output_tokens": [],
    }

    for token_type in ["total_tokens", "input_tokens", "output_tokens"]:
        running_total = 0
        for count in individual_usage[token_type]:
            if count is not None:
                running_total += count
            per_message_cumulative[token_type].append(running_total)

    return per_message_cumulative

find_rag_agent(mode)

Find the rag_agent with the given mode.

Source code in biochatter/llm_connect/conversation.py
def find_rag_agent(self, mode: str) -> tuple[int, RagAgent]:
    """Find the rag_agent with the given mode."""
    for i, val in enumerate(self.rag_agents):
        if val.mode == mode:
            return i, val
    return -1, None

get_last_injected_context()

Get a formatted list of the last context.

Get the last context injected into the conversation. Contains one dictionary for each RAG mode.

Returns
List[dict]: A list of dictionaries containing the mode and context
for each RAG agent.
Source code in biochatter/llm_connect/conversation.py
def get_last_injected_context(self) -> list[dict]:
    """Get a formatted list of the last context.

    Get the last context injected into the conversation. Contains one
    dictionary for each RAG mode.

    Returns
    -------
        List[dict]: A list of dictionaries containing the mode and context
        for each RAG agent.

    """
    return [{"mode": agent.mode, "context": agent.last_response} for agent in self.rag_agents]

get_msg_json()

Return a JSON representation of the conversation.

Returns a list of dicts of the messages in the conversation in JSON format. The keys of the dicts are the roles, the values are the messages.

Returns
str: A JSON representation of the messages in the conversation.
Source code in biochatter/llm_connect/conversation.py
def get_msg_json(self) -> str:
    """Return a JSON representation of the conversation.

    Returns a list of dicts of the messages in the conversation in JSON
    format. The keys of the dicts are the roles, the values are the
    messages.

    Returns
    -------
        str: A JSON representation of the messages in the conversation.

    """
    d = []
    for msg in self.messages:
        if isinstance(msg, SystemMessage):
            role = "system"
        elif isinstance(msg, HumanMessage):
            role = "user"
        elif isinstance(msg, AIMessage):
            role = "ai"
        else:
            error_msg = f"Unknown message type: {type(msg)}"
            raise TypeError(error_msg)

        d.append({role: msg.content})

    return json.dumps(d)

get_prompts()

Get the prompts.

Source code in biochatter/llm_connect/conversation.py
def get_prompts(self) -> dict:
    """Get the prompts."""
    return self.prompts

query(text, image_url=None, structured_model=None, wrap_structured_output=None, tools=None, explain_tool_result=None, additional_tools_instructions=None, general_instructions_tool_interpretation=None, additional_instructions_tool_interpretation=None, mcp=None, return_tool_calls_as_ai_message=None, track_tool_calls=None, **kwargs)

Query the LLM API using the user's query.

Appends the most recent query to the conversation, optionally injects context from the RAG agent, and runs the primary query method of the child class.


text (str): The user query.

image_url (str): The URL of an image to include in the conversation.
    Optional and only supported for models with vision capabilities.

structured_model (BaseModel): The structured output model to use for the query.

wrap_structured_output (bool): Whether to wrap the structured output in JSON quotes.

tools (list[Callable]): The tools to use for the query.

explain_tool_result (bool): Whether to explain the tool result.

additional_tools_instructions (str): The additional instructions for the query.
    Mainly used for tools that do not support tool calling.

general_instructions_tool_interpretation (str): The general
    instructions for the tool interpretation.
    Overrides the default prompt in `GENERAL_TOOL_RESULT_INTERPRETATION_PROMPT`.

additional_instructions_tool_interpretation (str): The additional
    instructions for the tool interpretation.
    Overrides the default prompt in `ADDITIONAL_TOOL_RESULT_INTERPRETATION_PROMPT`.

mcp (bool): If you want to use MCP mode, this should be set to True.

return_tool_calls_as_ai_message (bool): If you want to return the tool calls as an AI message, this should be set to True.

track_tool_calls (bool): If you want to track the tool calls, this should be set to True.

**kwargs: Additional keyword arguments.

tuple: A tuple containing the response from the API, the token usage
    information, and the correction if necessary/desired.
Source code in biochatter/llm_connect/conversation.py
def query(
    self,
    text: str,
    image_url: str | None = None,
    structured_model: BaseModel | None = None,
    wrap_structured_output: bool | None = None,
    tools: list[Callable] | None = None,
    explain_tool_result: bool | None = None,
    additional_tools_instructions: str | None = None,
    general_instructions_tool_interpretation: str | None = None,
    additional_instructions_tool_interpretation: str | None = None,
    mcp: bool | None = None,
    return_tool_calls_as_ai_message: bool | None = None,
    track_tool_calls: bool | None = None,
    **kwargs,
) -> tuple[str, dict | None, str | None]:
    """Query the LLM API using the user's query.

    Appends the most recent query to the conversation, optionally injects
    context from the RAG agent, and runs the primary query method of the
    child class.

    Args:
    ----
        text (str): The user query.

        image_url (str): The URL of an image to include in the conversation.
            Optional and only supported for models with vision capabilities.

        structured_model (BaseModel): The structured output model to use for the query.

        wrap_structured_output (bool): Whether to wrap the structured output in JSON quotes.

        tools (list[Callable]): The tools to use for the query.

        explain_tool_result (bool): Whether to explain the tool result.

        additional_tools_instructions (str): The additional instructions for the query.
            Mainly used for tools that do not support tool calling.

        general_instructions_tool_interpretation (str): The general
            instructions for the tool interpretation.
            Overrides the default prompt in `GENERAL_TOOL_RESULT_INTERPRETATION_PROMPT`.

        additional_instructions_tool_interpretation (str): The additional
            instructions for the tool interpretation.
            Overrides the default prompt in `ADDITIONAL_TOOL_RESULT_INTERPRETATION_PROMPT`.

        mcp (bool): If you want to use MCP mode, this should be set to True.

        return_tool_calls_as_ai_message (bool): If you want to return the tool calls as an AI message, this should be set to True.

        track_tool_calls (bool): If you want to track the tool calls, this should be set to True.

        **kwargs: Additional keyword arguments.

    Returns:
    -------
        tuple: A tuple containing the response from the API, the token usage
            information, and the correction if necessary/desired.

    """
    if mcp:
        self.mcp = True

    # save the last human prompt that may be used for answer enhancement
    self.last_human_prompt = text

    # if additional_tools_instructions are provided, save them
    if additional_tools_instructions:
        self.additional_tools_instructions = additional_tools_instructions

    # override the default prompts if other provided
    self.general_instructions_tool_interpretation = (
        general_instructions_tool_interpretation
        if general_instructions_tool_interpretation
        else GENERAL_TOOL_RESULT_INTERPRETATION_PROMPT
    )
    self.additional_instructions_tool_interpretation = (
        additional_instructions_tool_interpretation
        if additional_instructions_tool_interpretation
        else ADDITIONAL_TOOL_RESULT_INTERPRETATION_PROMPT
    )
    if not image_url:
        self.append_user_message(text)
    else:
        self.append_image_message(text, image_url)

    self._inject_context(text)

    # tools passed at this step are used only for this message
    msg, token_usage = self._primary_query(
        tools=tools,
        explain_tool_result=explain_tool_result,
        return_tool_calls_as_ai_message=return_tool_calls_as_ai_message,
        structured_model=structured_model,
        wrap_structured_output=wrap_structured_output,
        track_tool_calls=track_tool_calls,
    )

    # case of structured output
    if (token_usage == -1) and structured_model:
        return (msg, 0, None)

    if not token_usage:
        # indicates error
        return (msg, None, None)

    if not self.correct:
        return (msg, token_usage, None)

    cor_msg = "Correcting (using single sentences) ..." if self.split_correction else "Correcting ..."

    if st:
        with st.spinner(cor_msg):
            corrections = self._correct_query(text)
    else:
        corrections = self._correct_query(text)

    if not corrections:
        return (msg, token_usage, None)

    correction = "\n".join(corrections)
    return (msg, token_usage, correction)

reset()

Reset the conversation to the initial state.

Source code in biochatter/llm_connect/conversation.py
def reset(self) -> None:
    """Reset the conversation to the initial state."""
    self.history = []
    self.messages = []
    self.ca_messages = []
    self.current_statements = []
    self.tool_calls.clear()

set_api_key(api_key, user=None) abstractmethod

Set the API key.

Source code in biochatter/llm_connect/conversation.py
@abstractmethod
def set_api_key(self, api_key: str, user: str | None = None) -> None:
    """Set the API key."""

set_prompts(prompts)

Set the prompts.

Source code in biochatter/llm_connect/conversation.py
def set_prompts(self, prompts: dict) -> None:
    """Set the prompts."""
    self.prompts = prompts

set_rag_agent(agent)

Update or insert rag_agent.

If the rag_agent with the same mode already exists, it will be updated. Otherwise, the new rag_agent will be inserted.

Source code in biochatter/llm_connect/conversation.py
def set_rag_agent(self, agent: RagAgent) -> None:
    """Update or insert rag_agent.

    If the rag_agent with the same mode already exists, it will be updated.
    Otherwise, the new rag_agent will be inserted.
    """
    i, _ = self.find_rag_agent(agent.mode)
    if i < 0:
        # insert
        self.rag_agents.append(agent)
    else:
        # update
        self.rag_agents[i] = agent

set_user_name(user_name)

Set the user name.

Source code in biochatter/llm_connect/conversation.py
def set_user_name(self, user_name: str) -> None:
    """Set the user name."""
    self.user_name = user_name

setup(context)

Set up the conversation with general prompts and a context.

Source code in biochatter/llm_connect/conversation.py
def setup(self, context: str) -> None:
    """Set up the conversation with general prompts and a context."""
    for msg in self.prompts["primary_model_prompts"]:
        if msg:
            self.append_system_message(msg)

    for msg in self.prompts["correcting_agent_prompts"]:
        if msg:
            self.append_ca_message(msg)

    self.context = context
    msg = f"The topic of the research is {context}."
    self.append_system_message(msg)

setup_data_input_manual(data_input)

Set up the data input manually.

Source code in biochatter/llm_connect/conversation.py
def setup_data_input_manual(self, data_input: str) -> None:
    """Set up the data input manually."""
    self.data_input = data_input
    msg = f"The user has given information on the data input: {data_input}."
    self.append_system_message(msg)

setup_data_input_tool(df, input_file_name)

Set up the data input tool.

Source code in biochatter/llm_connect/conversation.py
def setup_data_input_tool(self, df, input_file_name: str) -> None:
    """Set up the data input tool."""
    self.data_input_tool = df

    for tool_name in self.prompts["tool_prompts"]:
        if tool_name in input_file_name:
            msg = self.prompts["tool_prompts"][tool_name].format(df=df)
            self.append_system_message(msg)

GeminiConversation

Bases: Conversation

Conversation class for the Google Gemini model.

Source code in biochatter/llm_connect/gemini.py
class GeminiConversation(Conversation):
    """Conversation class for the Google Gemini model."""

    def __init__(
        self,
        model_name: str,
        prompts: dict,
        correct: bool = False,
        split_correction: bool = False,
        tools: list[Callable] = None,
        tool_call_mode: Literal["auto", "text"] = "auto",
    ) -> None:
        """Initialise the GeminiConversation class.

        Connect to Google's Gemini API and set up a conversation with the user.
        Also initialise a second conversational agent to provide corrections to
        the model output, if necessary.

        Args:
        ----
            model_name (str): The name of the model to use.

            prompts (dict): A dictionary of prompts to use for the conversation.

            correct (bool): Whether to correct the model output.

            split_correction (bool): Whether to correct the model output by
                splitting the output into sentences and correcting each
                sentence individually.

            tools (list[Callable]): List of tool functions to use with the model.

            tool_call_mode (str): The mode to use for tool calls.
                "auto": Automatically call tools.
                "text": Only return text output of the tool call.

        """
        super().__init__(
            model_name=model_name,
            prompts=prompts,
            correct=correct,
            split_correction=split_correction,
            tools=tools,
            tool_call_mode=tool_call_mode,
        )

        self.ca_model_name = "gemini-2.0-flash"

    def set_api_key(self, api_key: str, user: str | None = None) -> bool:
        """Set the API key for the Google Gemini API.

        If the key is valid, initialise the conversational agent. Optionally set
        the user for usage statistics.

        Args:
        ----
            api_key (str): The API key for the Google Gemini API.

            user (str, optional): The user for usage statistics. If provided and
                equals "community", will track usage stats.

        Returns:
        -------
            bool: True if the API key is valid, False otherwise.

        """
        self.user = user

        try:
            self.chat = ChatGoogleGenerativeAI(
                model=self.model_name,
                temperature=0,
                google_api_key=api_key,
            )
            self.ca_chat = ChatGoogleGenerativeAI(
                model=self.ca_model_name,
                temperature=0,
                google_api_key=api_key,
            )

            # if binding happens here, tools will be available for all messages
            if self.tools:
                self.bind_tools(self.tools)

            return True

        except Exception:  # Google Genai doesn't expose specific exception types
            self._chat = None
            self._ca_chat = None
            return False

    def _primary_query(self, tools: list[Callable] | None = None, **kwargs) -> tuple:
        """Query the Google Gemini API with the user's message.

        Return the response using the message history (flattery system messages,
        prior conversation) as context. Correct the response if necessary.

        Args:
        ----
            tools (list[Callable]): The tools to use for the query. Tools
            passed at this step are used only for this message and not stored
            as part of the conversation object.

            **kwargs: Additional keyword arguments.

        Returns:
        -------
            tuple: A tuple containing the response from the Gemini API and
                the token usage.

        """
        if kwargs:
            kwargs.pop("tools", None)
            warnings.warn(f"Warning: {kwargs} are not used by this class", UserWarning)

        # bind tools to the chat if provided in the query
        chat = self.chat.bind_tools(tools) if (tools and self.model_name in TOOL_CALLING_MODELS) else self.chat

        try:
            response = chat.invoke(self.messages)
        except Exception as e:
            return str(e), None

        # Process tool calls if present
        if response.tool_calls:
            msg = self._process_tool_calls(response.tool_calls, tools, response.content)
        else:
            msg = response.content
            self.append_ai_message(msg)

        token_usage_raw = response.usage_metadata
        token_usage = self._extract_total_tokens(token_usage_raw)

        return msg, token_usage

    def _correct_response(self, msg: str) -> str:
        """Correct the response from the Gemini API.

        Send the response to a secondary language model. Optionally split the
        response into single sentences and correct each sentence individually.
        Update usage stats.

        Args:
        ----
            msg (str): The response from the Gemini API.

        Returns:
        -------
            str: The corrected response (or OK if no correction necessary).

        """
        ca_messages = self.ca_messages.copy()

        ca_messages.append(
            SystemMessage(
                content="If there is nothing to correct, please respond with just 'OK', and nothing else!",
            ),
        )

        ca_messages.append(
            HumanMessage(
                content=msg,
            ),
        )

        response = self.ca_chat.invoke(ca_messages)

        correction = response.content
        token_usage_raw = response.usage_metadata
        token_usage = self._extract_total_tokens(token_usage_raw)

        return correction

__init__(model_name, prompts, correct=False, split_correction=False, tools=None, tool_call_mode='auto')

Initialise the GeminiConversation class.

Connect to Google's Gemini API and set up a conversation with the user. Also initialise a second conversational agent to provide corrections to the model output, if necessary.


model_name (str): The name of the model to use.

prompts (dict): A dictionary of prompts to use for the conversation.

correct (bool): Whether to correct the model output.

split_correction (bool): Whether to correct the model output by
    splitting the output into sentences and correcting each
    sentence individually.

tools (list[Callable]): List of tool functions to use with the model.

tool_call_mode (str): The mode to use for tool calls.
    "auto": Automatically call tools.
    "text": Only return text output of the tool call.
Source code in biochatter/llm_connect/gemini.py
def __init__(
    self,
    model_name: str,
    prompts: dict,
    correct: bool = False,
    split_correction: bool = False,
    tools: list[Callable] = None,
    tool_call_mode: Literal["auto", "text"] = "auto",
) -> None:
    """Initialise the GeminiConversation class.

    Connect to Google's Gemini API and set up a conversation with the user.
    Also initialise a second conversational agent to provide corrections to
    the model output, if necessary.

    Args:
    ----
        model_name (str): The name of the model to use.

        prompts (dict): A dictionary of prompts to use for the conversation.

        correct (bool): Whether to correct the model output.

        split_correction (bool): Whether to correct the model output by
            splitting the output into sentences and correcting each
            sentence individually.

        tools (list[Callable]): List of tool functions to use with the model.

        tool_call_mode (str): The mode to use for tool calls.
            "auto": Automatically call tools.
            "text": Only return text output of the tool call.

    """
    super().__init__(
        model_name=model_name,
        prompts=prompts,
        correct=correct,
        split_correction=split_correction,
        tools=tools,
        tool_call_mode=tool_call_mode,
    )

    self.ca_model_name = "gemini-2.0-flash"

set_api_key(api_key, user=None)

Set the API key for the Google Gemini API.

If the key is valid, initialise the conversational agent. Optionally set the user for usage statistics.


api_key (str): The API key for the Google Gemini API.

user (str, optional): The user for usage statistics. If provided and
    equals "community", will track usage stats.

bool: True if the API key is valid, False otherwise.
Source code in biochatter/llm_connect/gemini.py
def set_api_key(self, api_key: str, user: str | None = None) -> bool:
    """Set the API key for the Google Gemini API.

    If the key is valid, initialise the conversational agent. Optionally set
    the user for usage statistics.

    Args:
    ----
        api_key (str): The API key for the Google Gemini API.

        user (str, optional): The user for usage statistics. If provided and
            equals "community", will track usage stats.

    Returns:
    -------
        bool: True if the API key is valid, False otherwise.

    """
    self.user = user

    try:
        self.chat = ChatGoogleGenerativeAI(
            model=self.model_name,
            temperature=0,
            google_api_key=api_key,
        )
        self.ca_chat = ChatGoogleGenerativeAI(
            model=self.ca_model_name,
            temperature=0,
            google_api_key=api_key,
        )

        # if binding happens here, tools will be available for all messages
        if self.tools:
            self.bind_tools(self.tools)

        return True

    except Exception:  # Google Genai doesn't expose specific exception types
        self._chat = None
        self._ca_chat = None
        return False

GptConversation

Bases: Conversation

Conversation class for the OpenAI GPT model.

Source code in biochatter/llm_connect/openai.py
class GptConversation(Conversation):
    """Conversation class for the OpenAI GPT model."""

    def __init__(
        self,
        model_name: str,
        prompts: dict,
        correct: bool = False,
        split_correction: bool = False,
        base_url: str = None,
        update_token_usage: Callable | None = None,
    ) -> None:
        """Connect to OpenAI's GPT API and set up a conversation with the user.

        Also initialise a second conversational agent to provide corrections to
        the model output, if necessary.

        Args:
        ----
            model_name (str): The name of the model to use.

            prompts (dict): A dictionary of prompts to use for the conversation.

            split_correction (bool): Whether to correct the model output by
                splitting the output into sentences and correcting each
                sentence individually.

            base_url (str): Optional OpenAI base_url value to use custom
                endpoint URL instead of default

        """
        super().__init__(
            model_name=model_name,
            prompts=prompts,
            correct=correct,
            split_correction=split_correction,
        )
        self.base_url = base_url
        self.ca_model_name = "gpt-3.5-turbo"
        # TODO make accessible by drop-down

        self._update_token_usage = update_token_usage

    def set_api_key(self, api_key: str, user: str | None = None) -> bool:
        """Set the API key for the OpenAI API.

        If the key is valid, initialise the conversational agent. Optionally set
        the user for usage statistics.

        Args:
        ----
            api_key (str): The API key for the OpenAI API.

            user (str, optional): The user for usage statistics. If provided and
                equals "community", will track usage stats.

        Returns:
        -------
            bool: True if the API key is valid, False otherwise.

        """
        client = openai.OpenAI(
            api_key=api_key,
            base_url=self.base_url,
        )
        self.user = user

        try:
            client.models.list()
            self.chat = ChatOpenAI(
                model_name=self.model_name,
                temperature=0,
                openai_api_key=api_key,
                base_url=self.base_url,
            )
            self.ca_chat = ChatOpenAI(
                model_name=self.ca_model_name,
                temperature=0,
                openai_api_key=api_key,
                base_url=self.base_url,
            )
            if user == "community":
                self.usage_stats = get_stats(user=user)

            return True

        except openai._exceptions.AuthenticationError:
            self._chat = None
            self._ca_chat = None
            return False

    def _primary_query(self, **kwargs) -> tuple:
        """Query the OpenAI API with the user's message.

        Return the response using the message history (flattery system messages,
        prior conversation) as context. Correct the response if necessary.

        Args:
        ----
            **kwargs: Keyword arguments (not used by this basic GPT implementation,
                     but accepted for compatibility with the base Conversation interface)

        Returns:
        -------
            tuple: A tuple containing the response from the OpenAI API and the
                token usage.

        """
        if kwargs:
            warnings.warn(f"Warning: {kwargs} are not used by this class", UserWarning)

        try:
            response = self.chat.generate([self.messages])
        except (
            openai._exceptions.APIError,
            openai._exceptions.OpenAIError,
            openai._exceptions.ConflictError,
            openai._exceptions.NotFoundError,
            openai._exceptions.APIStatusError,
            openai._exceptions.RateLimitError,
            openai._exceptions.APITimeoutError,
            openai._exceptions.BadRequestError,
            openai._exceptions.APIConnectionError,
            openai._exceptions.AuthenticationError,
            openai._exceptions.InternalServerError,
            openai._exceptions.PermissionDeniedError,
            openai._exceptions.UnprocessableEntityError,
            openai._exceptions.APIResponseValidationError,
        ) as e:
            return str(e), None

        msg = response.generations[0][0].text
        token_usage_raw = response.llm_output.get("token_usage")
        token_usage = self._extract_total_tokens(token_usage_raw)

        self._update_usage_stats(self.model_name, token_usage_raw)

        self.append_ai_message(msg)

        return msg, token_usage

    def _correct_response(self, msg: str) -> str:
        """Correct the response from the OpenAI API.

        Send the response to a secondary language model. Optionally split the
        response into single sentences and correct each sentence individually.
        Update usage stats.

        Args:
        ----
            msg (str): The response from the OpenAI API.

        Returns:
        -------
            str: The corrected response (or OK if no correction necessary).

        """
        ca_messages = self.ca_messages.copy()
        ca_messages.append(
            HumanMessage(
                content=msg,
            ),
        )
        ca_messages.append(
            SystemMessage(
                content="If there is nothing to correct, please respond with just 'OK', and nothing else!",
            ),
        )

        response = self.ca_chat.generate([ca_messages])

        correction = response.generations[0][0].text
        token_usage = response.llm_output.get("token_usage")

        self._update_usage_stats(self.ca_model_name, token_usage)

        return correction

    def _update_usage_stats(self, model: str, token_usage: dict) -> None:
        """Update redis database with token usage statistics.

        Use the usage_stats object with the increment method.

        Args:
        ----
            model (str): The model name.

            token_usage (dict): The token usage statistics.

        """
        if self.user == "community":
            # Only process integer values
            stats_dict = {f"{k}:{model}": v for k, v in token_usage.items() if isinstance(v, int | float)}
            self.usage_stats.increment(
                "usage:[date]:[user]",
                stats_dict,
            )

        if self._update_token_usage is not None:
            self._update_token_usage(self.user, model, token_usage)

__init__(model_name, prompts, correct=False, split_correction=False, base_url=None, update_token_usage=None)

Connect to OpenAI's GPT API and set up a conversation with the user.

Also initialise a second conversational agent to provide corrections to the model output, if necessary.


model_name (str): The name of the model to use.

prompts (dict): A dictionary of prompts to use for the conversation.

split_correction (bool): Whether to correct the model output by
    splitting the output into sentences and correcting each
    sentence individually.

base_url (str): Optional OpenAI base_url value to use custom
    endpoint URL instead of default
Source code in biochatter/llm_connect/openai.py
def __init__(
    self,
    model_name: str,
    prompts: dict,
    correct: bool = False,
    split_correction: bool = False,
    base_url: str = None,
    update_token_usage: Callable | None = None,
) -> None:
    """Connect to OpenAI's GPT API and set up a conversation with the user.

    Also initialise a second conversational agent to provide corrections to
    the model output, if necessary.

    Args:
    ----
        model_name (str): The name of the model to use.

        prompts (dict): A dictionary of prompts to use for the conversation.

        split_correction (bool): Whether to correct the model output by
            splitting the output into sentences and correcting each
            sentence individually.

        base_url (str): Optional OpenAI base_url value to use custom
            endpoint URL instead of default

    """
    super().__init__(
        model_name=model_name,
        prompts=prompts,
        correct=correct,
        split_correction=split_correction,
    )
    self.base_url = base_url
    self.ca_model_name = "gpt-3.5-turbo"
    # TODO make accessible by drop-down

    self._update_token_usage = update_token_usage

set_api_key(api_key, user=None)

Set the API key for the OpenAI API.

If the key is valid, initialise the conversational agent. Optionally set the user for usage statistics.


api_key (str): The API key for the OpenAI API.

user (str, optional): The user for usage statistics. If provided and
    equals "community", will track usage stats.

bool: True if the API key is valid, False otherwise.
Source code in biochatter/llm_connect/openai.py
def set_api_key(self, api_key: str, user: str | None = None) -> bool:
    """Set the API key for the OpenAI API.

    If the key is valid, initialise the conversational agent. Optionally set
    the user for usage statistics.

    Args:
    ----
        api_key (str): The API key for the OpenAI API.

        user (str, optional): The user for usage statistics. If provided and
            equals "community", will track usage stats.

    Returns:
    -------
        bool: True if the API key is valid, False otherwise.

    """
    client = openai.OpenAI(
        api_key=api_key,
        base_url=self.base_url,
    )
    self.user = user

    try:
        client.models.list()
        self.chat = ChatOpenAI(
            model_name=self.model_name,
            temperature=0,
            openai_api_key=api_key,
            base_url=self.base_url,
        )
        self.ca_chat = ChatOpenAI(
            model_name=self.ca_model_name,
            temperature=0,
            openai_api_key=api_key,
            base_url=self.base_url,
        )
        if user == "community":
            self.usage_stats = get_stats(user=user)

        return True

    except openai._exceptions.AuthenticationError:
        self._chat = None
        self._ca_chat = None
        return False

LangChainConversation

Bases: Conversation

Conversation class for a generic LangChain model.

Source code in biochatter/llm_connect/langchain.py
class LangChainConversation(Conversation):
    """Conversation class for a generic LangChain model."""

    def __init__(
        self,
        model_name: str,
        model_provider: str,
        prompts: dict,
        correct: bool = False,
        split_correction: bool = False,
        tools: list[Callable] = None,
        tool_call_mode: Literal["auto", "text"] = "auto",
        async_mode: bool = False,
        mcp: bool = False,
        force_tool: bool = False,
    ) -> None:
        """Initialise the LangChainConversation class.

        Connect to a generic LangChain model and set up a conversation with the
        user. Also initialise a second conversational agent to provide
        corrections to the model output, if necessary.

        Args:
        ----
            model_name (str): The name of the model to use.
            model_provider (str): The provider of the model to use.
            prompts (dict): A dictionary of prompts to use for the conversation.
            correct (bool): Whether to correct the model output.
            split_correction (bool): Whether to correct the model output by
                splitting the output into sentences and correcting each
                sentence individually.
            tools (list[Callable]): List of tool functions to use with the
                model.
            tool_call_mode (str): The mode to use for tool calls.
                "auto": Automatically call tools.
                "text": Only return text output of the tool call.
            async_mode (bool): Whether to run in async mode. Defaults to False.
            mcp (bool): If you want to use MCP mode, this should be set to True.
            force_tool (bool): If you want to force the model to use tools, this should be set to True.

        """
        super().__init__(
            model_name=model_name,
            prompts=prompts,
            correct=correct,
            split_correction=split_correction,
            tools=tools,
            tool_call_mode=tool_call_mode,
            mcp=mcp,
            force_tool=force_tool,
        )

        self.model_name = model_name
        self.model_provider = model_provider
        self.async_mode = async_mode

    # TODO: the name of this method is overloaded, since the api key is loaded
    # from the environment variables and not as an argument
    def set_api_key(self, api_key: str | None = None, user: str | None = None) -> bool:
        """Set the API key for the model provider.

        If the key is valid, initialise the conversational agent. Optionally set
        the user for usage statistics.

        Args:
        ----
            api_key (str): The API key for the model provider.

            user (str, optional): The user for usage statistics. If provided and
                equals "community", will track usage stats.

        Returns:
        -------
            bool: True if the API key is valid, False otherwise.

        """
        self.user = user

        try:
            self.chat = init_chat_model(
                model=self.model_name,
                model_provider=self.model_provider,
                temperature=0,
            )
            self.ca_chat = init_chat_model(
                model=self.model_name,
                model_provider=self.model_provider,
                temperature=0,
            )

            # if binding happens here, tools will be available for all messages
            if self.tools:
                self.bind_tools(self.tools)

            return True

        except Exception:  # Google Genai doesn't expose specific exception types
            self._chat = None
            self._ca_chat = None
            return False

    def _primary_query(
        self,
        tools: list[Callable] | None = None,
        explain_tool_result: bool = False,
        return_tool_calls_as_ai_message: bool = False,
        structured_model: BaseModel | None = None,
        wrap_structured_output: bool = False,
        track_tool_calls: bool = False,
    ) -> tuple:
        """Run the primary query.

        Args:
        ----
            tools (list[Callable], optional): Additional tools to use for this specific query.
            explain_tool_result (bool, optional): Whether to explain the tool result.
            return_tool_calls_as_ai_message (bool, optional): Whether to return tool calls as an AI message.
            structured_model (BaseModel, optional): The structured output model to use.
            wrap_structured_output (bool, optional): Whether to wrap the structured output in JSON quotes.
            track_tool_calls (bool, optional): Whether to track the tool calls.

        Returns:
        -------
            tuple: A tuple containing the response message and token usage information.

        """
        token_usage = None  # Initialize token_usage
        msg = None  # Initialize msg

        starting_tools = self.tools if self.tools else []
        in_chat_tools = tools if tools else []
        available_tools = starting_tools + in_chat_tools

        if structured_model and len(available_tools) > 0:
            raise ValueError("Structured output and tools cannot be used together at the moment.")

        if self.model_name in STRUCTURED_OUTPUT_MODELS and structured_model:
            chat = self.chat.with_structured_output(structured_model)
        elif structured_model and self.model_name not in STRUCTURED_OUTPUT_MODELS:
            # add to the end of the prompt an instruction to return a structured output
            chat = self.chat
            self.messages[-1].content = (
                self.messages[-1].content
                + "\n\nPlease return a structured output following this schema: "
                + str(structured_model.model_json_schema())
                + (
                    " Just return the JSON object wrapped in ```json tags and nothing else."
                    if wrap_structured_output
                    else " Just return the JSON object and nothing else."
                )
            )

        if (self.model_name in TOOL_CALLING_MODELS and not structured_model) or self.force_tool:
            chat = self.chat.bind_tools(available_tools)
        elif self.model_name not in TOOL_CALLING_MODELS and len(available_tools) > 0:
            self.tools_prompt = self._create_tool_prompt(
                tools=available_tools,
                additional_instructions=self.additional_tools_instructions,
            )
            if not self.messages:
                msg = "No messages available in the conversation"
                raise ValueError(msg)
            self.messages[-1] = self.tools_prompt
            chat = self.chat
        elif len(available_tools) == 0 and not structured_model:
            chat = self.chat

        try:
            response = chat.invoke(self.messages)
        except Exception as e:
            return str(e), None

        # Structured output don't have tool calls attribute
        if hasattr(response, "tool_calls"):
            token_usage_raw = response.usage_metadata if response.usage_metadata else None
            token_usage = self._extract_total_tokens(token_usage_raw)
            # case in which the model called tools
            if len(response.tool_calls) > 0:
                self.append_ai_message(response)
                msg = self._process_tool_calls(
                    tool_calls=response.tool_calls,
                    available_tools=available_tools,
                    response_content=response.content,
                    explain_tool_result=explain_tool_result,
                    return_tool_calls_as_ai_message=return_tool_calls_as_ai_message,
                    track_tool_calls=track_tool_calls,
                )
            # case where the model does not support tool calling natively, called a tool and we need manual processing
            elif self.model_name not in TOOL_CALLING_MODELS and self.tools_prompt:
                cleaned_content = (
                    response.content.replace('"""', "").replace("json", "").replace("`", "").replace("\n", "").strip()
                )
                try:
                    tool_call_data = json.loads(cleaned_content)
                    msg = self._process_manual_tool_call(
                        tool_call=tool_call_data,
                        available_tools=available_tools,
                        explain_tool_result=explain_tool_result,
                    )
                    # token_usage remains None (from line 176) as per successful tool call path logic
                except json.JSONDecodeError:
                    # If JSON parsing fails, the model didn't return a valid tool call.
                    # Treat as a regular message from the LLM.
                    msg = response.content  # Use original content
                    # Update token_usage, similar to 'no tool calls' or 'manual structured output' paths
            # case where the model does not support structured output but the user has provided a structured model
            elif self.model_name not in STRUCTURED_OUTPUT_MODELS and structured_model:
                # check that the output conforms to the structured model
                pydantic_manual_validator(response.content, structured_model)
                msg = response.content

            # no tool calls
            else:
                msg = response.content
                self.append_ai_message(response)

        # even if there are no tool calls, the standard langchain output has a tool_calls attribute
        # therefore, this case only happens when the returned ouput from the invoke is a structured output
        else:
            msg = response.model_dump_json()
            if wrap_structured_output:
                msg = "```json\n" + msg + "\n```"
            # we don't return None because is used to signal an error, instead here we just can't count the tokens
            token_usage = -1
            self.append_ai_message(msg)

        return msg, token_usage

    def _correct_response(self, msg: str) -> str:
        """Correct the response from the Gemini API.

        Send the response to a secondary language model. Optionally split the
        response into single sentences and correct each sentence individually.
        Update usage stats.

        Args:
        ----
            msg (str): The response from the Gemini API.

        Returns:
        -------
            str: The corrected response (or OK if no correction necessary).

        """
        ca_messages = self.ca_messages.copy()
        ca_messages.append(
            HumanMessage(
                content=msg,
            ),
        )
        ca_messages.append(
            SystemMessage(
                content="If there is nothing to correct, please respond with just 'OK', and nothing else!",
            ),
        )

        response = self.ca_chat.invoke(ca_messages)

        correction = response.content
        token_usage_raw = response.usage_metadata
        token_usage = self._extract_total_tokens(token_usage_raw)

        return correction, token_usage

__init__(model_name, model_provider, prompts, correct=False, split_correction=False, tools=None, tool_call_mode='auto', async_mode=False, mcp=False, force_tool=False)

Initialise the LangChainConversation class.

Connect to a generic LangChain model and set up a conversation with the user. Also initialise a second conversational agent to provide corrections to the model output, if necessary.


model_name (str): The name of the model to use.
model_provider (str): The provider of the model to use.
prompts (dict): A dictionary of prompts to use for the conversation.
correct (bool): Whether to correct the model output.
split_correction (bool): Whether to correct the model output by
    splitting the output into sentences and correcting each
    sentence individually.
tools (list[Callable]): List of tool functions to use with the
    model.
tool_call_mode (str): The mode to use for tool calls.
    "auto": Automatically call tools.
    "text": Only return text output of the tool call.
async_mode (bool): Whether to run in async mode. Defaults to False.
mcp (bool): If you want to use MCP mode, this should be set to True.
force_tool (bool): If you want to force the model to use tools, this should be set to True.
Source code in biochatter/llm_connect/langchain.py
def __init__(
    self,
    model_name: str,
    model_provider: str,
    prompts: dict,
    correct: bool = False,
    split_correction: bool = False,
    tools: list[Callable] = None,
    tool_call_mode: Literal["auto", "text"] = "auto",
    async_mode: bool = False,
    mcp: bool = False,
    force_tool: bool = False,
) -> None:
    """Initialise the LangChainConversation class.

    Connect to a generic LangChain model and set up a conversation with the
    user. Also initialise a second conversational agent to provide
    corrections to the model output, if necessary.

    Args:
    ----
        model_name (str): The name of the model to use.
        model_provider (str): The provider of the model to use.
        prompts (dict): A dictionary of prompts to use for the conversation.
        correct (bool): Whether to correct the model output.
        split_correction (bool): Whether to correct the model output by
            splitting the output into sentences and correcting each
            sentence individually.
        tools (list[Callable]): List of tool functions to use with the
            model.
        tool_call_mode (str): The mode to use for tool calls.
            "auto": Automatically call tools.
            "text": Only return text output of the tool call.
        async_mode (bool): Whether to run in async mode. Defaults to False.
        mcp (bool): If you want to use MCP mode, this should be set to True.
        force_tool (bool): If you want to force the model to use tools, this should be set to True.

    """
    super().__init__(
        model_name=model_name,
        prompts=prompts,
        correct=correct,
        split_correction=split_correction,
        tools=tools,
        tool_call_mode=tool_call_mode,
        mcp=mcp,
        force_tool=force_tool,
    )

    self.model_name = model_name
    self.model_provider = model_provider
    self.async_mode = async_mode

set_api_key(api_key=None, user=None)

Set the API key for the model provider.

If the key is valid, initialise the conversational agent. Optionally set the user for usage statistics.


api_key (str): The API key for the model provider.

user (str, optional): The user for usage statistics. If provided and
    equals "community", will track usage stats.

bool: True if the API key is valid, False otherwise.
Source code in biochatter/llm_connect/langchain.py
def set_api_key(self, api_key: str | None = None, user: str | None = None) -> bool:
    """Set the API key for the model provider.

    If the key is valid, initialise the conversational agent. Optionally set
    the user for usage statistics.

    Args:
    ----
        api_key (str): The API key for the model provider.

        user (str, optional): The user for usage statistics. If provided and
            equals "community", will track usage stats.

    Returns:
    -------
        bool: True if the API key is valid, False otherwise.

    """
    self.user = user

    try:
        self.chat = init_chat_model(
            model=self.model_name,
            model_provider=self.model_provider,
            temperature=0,
        )
        self.ca_chat = init_chat_model(
            model=self.model_name,
            model_provider=self.model_provider,
            temperature=0,
        )

        # if binding happens here, tools will be available for all messages
        if self.tools:
            self.bind_tools(self.tools)

        return True

    except Exception:  # Google Genai doesn't expose specific exception types
        self._chat = None
        self._ca_chat = None
        return False

LiteLLMConversation

Bases: Conversation

A unified interface for multiple LLM models using LiteLLM.

This class implements the abstract methods from the Conversation parent class and provides a unified way to interact with different LLM providers through LiteLLM, which supports models from OpenAI, Anthropic, HuggingFace, and more.

Attributes
model_name (str): The name of the model to use.
prompts (dict): Dictionary containing various prompts used in the conversation.
correct (bool): Whether to use a correcting agent.
split_correction (bool): Whether to split corrections by sentence.
rag_agents (list): List of RAG agents available for context enhancement.
history (list): Conversation history for logging/printing.
messages (list): Messages in the conversation.
ca_messages (list): Messages for the correcting agent.
api_key (str): API key for the LLM provider.
user (str): Username for the API, if required.
Source code in biochatter/llm_connect/llmlite.py
class LiteLLMConversation(Conversation):
    """A unified interface for multiple LLM models using LiteLLM.

    This class implements the abstract methods from the Conversation parent class
    and provides a unified way to interact with different LLM providers through
    LiteLLM, which supports models from OpenAI, Anthropic, HuggingFace, and more.

    Attributes
    ----------
        model_name (str): The name of the model to use.
        prompts (dict): Dictionary containing various prompts used in the conversation.
        correct (bool): Whether to use a correcting agent.
        split_correction (bool): Whether to split corrections by sentence.
        rag_agents (list): List of RAG agents available for context enhancement.
        history (list): Conversation history for logging/printing.
        messages (list): Messages in the conversation.
        ca_messages (list): Messages for the correcting agent.
        api_key (str): API key for the LLM provider.
        user (str): Username for the API, if required.

    """

    def __init__(
        self,
        model_name: str,
        prompts: dict,
        correct: bool = False,
        split_correction: bool = False,
        use_ragagent_selector: bool = False,
        update_token_usage: Callable | None = None,
    ) -> None:
        """Initialize a UnifiedConversation instance.

        Args:
        ----
            model_name (str): The name of the model to use.
            prompts (dict): Dictionary containing various prompts used in the conversation.
            correct (bool): Whether to use a correcting agent. Defaults to False.
            split_correction (bool): Whether to split corrections by sentence. Defaults to False.
            use_ragagent_selector (bool): Whether to use RagAgentSelector. Defaults to False.
            update_token_usage (Callable): A function to update the token usage statistics.

        """
        super().__init__(
            model_name=model_name,
            prompts=prompts,
            correct=correct,
            split_correction=split_correction,
            use_ragagent_selector=use_ragagent_selector,
        )
        self.api_key = None
        self.user = None
        self.ca_model_name = model_name
        self._update_token_usage = update_token_usage

    def get_litellm_object(self, api_key: str, model: str) -> ChatLiteLLM:
        """Get a LiteLLM object for the specified model and API key.

        Args:
        ----
            api_key (str): The API key for the LLM provider.
            model (str): The name of the model to use.

        Returns:
        -------
            ChatLiteLLM: An instance of ChatLiteLLM configured with the specified model, temperature, max tokens and API key.

        Raises:
        ------
            ValueError: If the API key is None.
            litellm.exceptions.AuthenticationError: If there is an authentication error.
            litellm.exceptions.InvalidRequestError: If the request is invalid.
            litellm.exceptions.RateLimitError: If the rate limit is exceeded.
            litellm.exceptions.ServiceUnavailableError: If the service is unavailable.
            litellm.exceptions.APIError: If there is a general API error.
            litellm.exceptions.Timeout: If the request times out.
            litellm.exceptions.APIConnectionError: If there is a connection error.
            litellm.exceptions.InternalServerError: If there is an internal server error.
            Exception: If there is an unexpected error.

        """
        if api_key is None:
            raise ValueError("API key must not be None")

        try:
            max_tokens = self.get_model_max_tokens(model)
        except:
            max_tokens = None

        kwargs = {"temperature": 0, "max_token": max_tokens, "model_name": model}

        if self.model_name.startswith("gpt-"):
            api_key_kwarg = "openai_api_key"
        elif self.model_name.startswith("claude-"):
            api_key_kwarg = "anthropic_api_key"
        elif self.model_name.startswith("azure/"):
            api_key_kwarg = "azure_api_key"
        elif self.model_name.startswith("mistral/") or self.model_name in [
            "mistral-tiny",
            "mistral-small",
            "mistral-medium",
            "mistral-large-latest",
        ]:
            api_key_kwarg = "api_key"
        else:
            api_key_kwarg = "api_key"

        kwargs[api_key_kwarg] = api_key
        try:
            return ChatLiteLLM(**kwargs)

        except (
            litellm.exceptions.AuthenticationError,
            litellm.exceptions.InvalidRequestError,
            litellm.exceptions.RateLimitError,
            litellm.exceptions.ServiceUnavailableError,
            litellm.exceptions.APIError,
            litellm.exceptions.Timeout,
            litellm.exceptions.APIConnectionError,
            litellm.exceptions.InternalServerError,
        ) as api_setup_error:
            raise api_setup_error

    def set_api_key(self, api_key: str, user: str | None = None) -> bool:
        """Set the API key for the LLM provider.

        Args:
        ----
            api_key (str): The API key for the LLM provider.
            user (Union[str, None]): The username

        Returns:
        -------
            bool: True if the API key is successfully set, False otherwise.

        Raises:
        ------
            ValueError: If the model name or correction model name is not set.
            TypeError: If the LiteLLM object initialization fails.
            Exception: If there is an unexpected error.

        """
        try:
            if self.model_name is None:
                raise ValueError("Primary Model name is not set.")

            if self.ca_model_name is None:
                raise ValueError("Correction Model name is not set.")

            self.chat = self.get_litellm_object(api_key, self.model_name)
            if self.chat is None:
                raise TypeError("Failed to intialize primary agent chat object.")

            self.ca_chat = self.get_litellm_object(api_key, self.ca_model_name)
            if self.ca_chat is None:
                raise TypeError("Failed to intialize correcting agent chat object.")

            self.user = user
            if user == "community":
                self.usage_stats = get_stats(user=user)
            return True

        except (ValueError, TypeError):
            self.chat = None
            self.ca_chat = None
            return False
        except Exception:
            self.chat = None
            self.ca_chat = None
            return False

    def json_serializable(self, obj):
        """Convert non-serializable objects to serializable format."""
        if obj is None:
            raise ValueError("Object is None")
        if hasattr(obj, "__dict__"):
            return obj.__dict__
        if hasattr(obj, "dict") and callable(obj.dict):
            return obj.dict()
        try:
            return str(obj)
        except:
            return repr(obj)

    def parse_llm_response(self, response) -> dict | None:
        """Parse the response from the LLM."""
        try:
            full_json = json.loads(json.dumps(response, default=self.json_serializable))

            if not full_json.get("generations"):
                return None

            generations = full_json["generations"]
            if not generations or not generations[0]:
                return None

            first_generation = generations[0][0]
            if not first_generation or not first_generation.get("message"):
                return None

            message = first_generation["message"]
            if not message.get("response_metadata"):
                return None

            response_metadata = message["response_metadata"]
            if not response_metadata.get("token_usage"):
                return None

            return response_metadata["token_usage"]

        except (KeyError, IndexError, TypeError, json.JSONDecodeError) as e:
            print(f"Error parsing LLM response: {e}")
            return None

        except Exception as e:
            print(f"Unexpected error while parsing LLM response: {e}")
            return None

    def _primary_query(self, **kwargs) -> tuple:
        """Query the LLM API with the user's message.

        Return the response using the message history (flattery system messages,
        prior conversation) as context. Correct the response if necessary.

        Args:
        ----
            **kwargs: Keyword arguments (not used by this basic LiteLLM implementation,
                     but accepted for compatibility with the base Conversation interface)

        Returns:
        -------
            tuple: A tuple containing the response from the LLM API and the token usage.

        """
        if kwargs:
            warnings.warn(f"Warning: {kwargs} are not used by this class", UserWarning)

        try:
            response = self.chat.generate([self.messages])
        except (
            AttributeError,
            litellm.exceptions.APIError,
            litellm.exceptions.OpenAIError,
            litellm.exceptions.RateLimitError,
            litellm.exceptions.APIConnectionError,
            litellm.exceptions.BadRequestError,
            litellm.exceptions.AuthenticationError,
            litellm.exceptions.InternalServerError,
            litellm.exceptions.PermissionDeniedError,
            litellm.exceptions.UnprocessableEntityError,
            litellm.exceptions.APIResponseValidationError,
            litellm.exceptions.BudgetExceededError,
            litellm.exceptions.RejectedRequestError,
            litellm.exceptions.ServiceUnavailableError,
            litellm.exceptions.Timeout,
        ) as e:
            return e, None
        except Exception as e:
            return e, None

        msg = response.generations[0][0].text
        token_usage_raw = self.parse_llm_response(response)
        token_usage = self._extract_total_tokens(token_usage_raw)

        self.append_ai_message(msg)

        self._update_usage_stats(self.model_name, token_usage)

        return msg, token_usage

    def _correct_response(self, msg: str) -> str:
        """Correct the response from the LLM.

        Args:
        ----
            msg (str): The response message to correct.

        Returns:
        -------
            str: The corrected response message.

        """
        ca_messages = self.ca_messages.copy()
        ca_messages.append(
            HumanMessage(
                content=msg,
            ),
        )
        ca_messages.append(
            SystemMessage(
                content="If there is nothing to correct, please respond with just 'OK', and nothing else!",
            ),
        )

        response = self.ca_chat.generate([ca_messages])

        correction = response.generations[0][0].text
        token_usage_raw = self.parse_llm_response(response)
        token_usage = self._extract_total_tokens(token_usage_raw)

        self._update_usage_stats(self.ca_model_name, token_usage_raw)

        return correction

    def _update_usage_stats(self, model: str, token_usage: dict) -> None:
        """Update the usage statistics.

        Args:
        ----
            model (str): The model name.
            token_usage (dict): The token usage information.

        """
        if self.user == "community" and model:
            stats_dict = {f"{k}:{model}": v for k, v in token_usage.items() if isinstance(v, int | float)}
            self.usage_stats.increment(
                "usage:[date]:[user]",
                stats_dict,
            )

        if self._update_token_usage is not None:
            self._update_token_usage(self.user, model, token_usage)

    def get_all_model_list(self) -> list:
        """Get a list of all available models."""
        return litellm.model_list

    def get_models_by_provider(self):
        """Get a dictionary of models grouped by their provider."""
        return litellm.models_by_provider

    def get_all_model_info(self) -> dict:
        """Get information about all available models."""
        return litellm.model_cost

    def get_model_info(self, model: str) -> dict:
        """Get information about a specific model.

        Args:
        ----
            model (str): The name of the model.

        Returns:
        -------
            dict: A dictionary containing information about the specified model.

        """
        models_info: dict = self.get_all_model_info()
        if model not in models_info:
            raise litellm.exceptions.NotFoundError(
                f"{model} model's information is not available.",
                model=model,
                llm_provider="Unknown",
            )
        return models_info[model]

    def get_model_max_tokens(self, model: str) -> int:
        """Get the maximum number of tokens for a specific model.

        Args:
        ----
            model (str): The name of the model.

        Returns:
        -------
            int: The maximum number of tokens for the specified model.

        """
        try:
            model_info = self.get_model_info(model)
            if "max_tokens" not in model_info:
                raise litellm.exceptions.NotFoundError(
                    f"Max token information for {model} is not available.",
                    model=model,
                    llm_provider="Unknown",
                )
            return model_info["max_tokens"]
        except litellm.exceptions.NotFoundError as e:
            raise e

__init__(model_name, prompts, correct=False, split_correction=False, use_ragagent_selector=False, update_token_usage=None)

Initialize a UnifiedConversation instance.


model_name (str): The name of the model to use.
prompts (dict): Dictionary containing various prompts used in the conversation.
correct (bool): Whether to use a correcting agent. Defaults to False.
split_correction (bool): Whether to split corrections by sentence. Defaults to False.
use_ragagent_selector (bool): Whether to use RagAgentSelector. Defaults to False.
update_token_usage (Callable): A function to update the token usage statistics.
Source code in biochatter/llm_connect/llmlite.py
def __init__(
    self,
    model_name: str,
    prompts: dict,
    correct: bool = False,
    split_correction: bool = False,
    use_ragagent_selector: bool = False,
    update_token_usage: Callable | None = None,
) -> None:
    """Initialize a UnifiedConversation instance.

    Args:
    ----
        model_name (str): The name of the model to use.
        prompts (dict): Dictionary containing various prompts used in the conversation.
        correct (bool): Whether to use a correcting agent. Defaults to False.
        split_correction (bool): Whether to split corrections by sentence. Defaults to False.
        use_ragagent_selector (bool): Whether to use RagAgentSelector. Defaults to False.
        update_token_usage (Callable): A function to update the token usage statistics.

    """
    super().__init__(
        model_name=model_name,
        prompts=prompts,
        correct=correct,
        split_correction=split_correction,
        use_ragagent_selector=use_ragagent_selector,
    )
    self.api_key = None
    self.user = None
    self.ca_model_name = model_name
    self._update_token_usage = update_token_usage

get_all_model_info()

Get information about all available models.

Source code in biochatter/llm_connect/llmlite.py
def get_all_model_info(self) -> dict:
    """Get information about all available models."""
    return litellm.model_cost

get_all_model_list()

Get a list of all available models.

Source code in biochatter/llm_connect/llmlite.py
def get_all_model_list(self) -> list:
    """Get a list of all available models."""
    return litellm.model_list

get_litellm_object(api_key, model)

Get a LiteLLM object for the specified model and API key.


api_key (str): The API key for the LLM provider.
model (str): The name of the model to use.

ChatLiteLLM: An instance of ChatLiteLLM configured with the specified model, temperature, max tokens and API key.

ValueError: If the API key is None.
litellm.exceptions.AuthenticationError: If there is an authentication error.
litellm.exceptions.InvalidRequestError: If the request is invalid.
litellm.exceptions.RateLimitError: If the rate limit is exceeded.
litellm.exceptions.ServiceUnavailableError: If the service is unavailable.
litellm.exceptions.APIError: If there is a general API error.
litellm.exceptions.Timeout: If the request times out.
litellm.exceptions.APIConnectionError: If there is a connection error.
litellm.exceptions.InternalServerError: If there is an internal server error.
Exception: If there is an unexpected error.
Source code in biochatter/llm_connect/llmlite.py
def get_litellm_object(self, api_key: str, model: str) -> ChatLiteLLM:
    """Get a LiteLLM object for the specified model and API key.

    Args:
    ----
        api_key (str): The API key for the LLM provider.
        model (str): The name of the model to use.

    Returns:
    -------
        ChatLiteLLM: An instance of ChatLiteLLM configured with the specified model, temperature, max tokens and API key.

    Raises:
    ------
        ValueError: If the API key is None.
        litellm.exceptions.AuthenticationError: If there is an authentication error.
        litellm.exceptions.InvalidRequestError: If the request is invalid.
        litellm.exceptions.RateLimitError: If the rate limit is exceeded.
        litellm.exceptions.ServiceUnavailableError: If the service is unavailable.
        litellm.exceptions.APIError: If there is a general API error.
        litellm.exceptions.Timeout: If the request times out.
        litellm.exceptions.APIConnectionError: If there is a connection error.
        litellm.exceptions.InternalServerError: If there is an internal server error.
        Exception: If there is an unexpected error.

    """
    if api_key is None:
        raise ValueError("API key must not be None")

    try:
        max_tokens = self.get_model_max_tokens(model)
    except:
        max_tokens = None

    kwargs = {"temperature": 0, "max_token": max_tokens, "model_name": model}

    if self.model_name.startswith("gpt-"):
        api_key_kwarg = "openai_api_key"
    elif self.model_name.startswith("claude-"):
        api_key_kwarg = "anthropic_api_key"
    elif self.model_name.startswith("azure/"):
        api_key_kwarg = "azure_api_key"
    elif self.model_name.startswith("mistral/") or self.model_name in [
        "mistral-tiny",
        "mistral-small",
        "mistral-medium",
        "mistral-large-latest",
    ]:
        api_key_kwarg = "api_key"
    else:
        api_key_kwarg = "api_key"

    kwargs[api_key_kwarg] = api_key
    try:
        return ChatLiteLLM(**kwargs)

    except (
        litellm.exceptions.AuthenticationError,
        litellm.exceptions.InvalidRequestError,
        litellm.exceptions.RateLimitError,
        litellm.exceptions.ServiceUnavailableError,
        litellm.exceptions.APIError,
        litellm.exceptions.Timeout,
        litellm.exceptions.APIConnectionError,
        litellm.exceptions.InternalServerError,
    ) as api_setup_error:
        raise api_setup_error

get_model_info(model)

Get information about a specific model.


model (str): The name of the model.

dict: A dictionary containing information about the specified model.
Source code in biochatter/llm_connect/llmlite.py
def get_model_info(self, model: str) -> dict:
    """Get information about a specific model.

    Args:
    ----
        model (str): The name of the model.

    Returns:
    -------
        dict: A dictionary containing information about the specified model.

    """
    models_info: dict = self.get_all_model_info()
    if model not in models_info:
        raise litellm.exceptions.NotFoundError(
            f"{model} model's information is not available.",
            model=model,
            llm_provider="Unknown",
        )
    return models_info[model]

get_model_max_tokens(model)

Get the maximum number of tokens for a specific model.


model (str): The name of the model.

int: The maximum number of tokens for the specified model.
Source code in biochatter/llm_connect/llmlite.py
def get_model_max_tokens(self, model: str) -> int:
    """Get the maximum number of tokens for a specific model.

    Args:
    ----
        model (str): The name of the model.

    Returns:
    -------
        int: The maximum number of tokens for the specified model.

    """
    try:
        model_info = self.get_model_info(model)
        if "max_tokens" not in model_info:
            raise litellm.exceptions.NotFoundError(
                f"Max token information for {model} is not available.",
                model=model,
                llm_provider="Unknown",
            )
        return model_info["max_tokens"]
    except litellm.exceptions.NotFoundError as e:
        raise e

get_models_by_provider()

Get a dictionary of models grouped by their provider.

Source code in biochatter/llm_connect/llmlite.py
def get_models_by_provider(self):
    """Get a dictionary of models grouped by their provider."""
    return litellm.models_by_provider

json_serializable(obj)

Convert non-serializable objects to serializable format.

Source code in biochatter/llm_connect/llmlite.py
def json_serializable(self, obj):
    """Convert non-serializable objects to serializable format."""
    if obj is None:
        raise ValueError("Object is None")
    if hasattr(obj, "__dict__"):
        return obj.__dict__
    if hasattr(obj, "dict") and callable(obj.dict):
        return obj.dict()
    try:
        return str(obj)
    except:
        return repr(obj)

parse_llm_response(response)

Parse the response from the LLM.

Source code in biochatter/llm_connect/llmlite.py
def parse_llm_response(self, response) -> dict | None:
    """Parse the response from the LLM."""
    try:
        full_json = json.loads(json.dumps(response, default=self.json_serializable))

        if not full_json.get("generations"):
            return None

        generations = full_json["generations"]
        if not generations or not generations[0]:
            return None

        first_generation = generations[0][0]
        if not first_generation or not first_generation.get("message"):
            return None

        message = first_generation["message"]
        if not message.get("response_metadata"):
            return None

        response_metadata = message["response_metadata"]
        if not response_metadata.get("token_usage"):
            return None

        return response_metadata["token_usage"]

    except (KeyError, IndexError, TypeError, json.JSONDecodeError) as e:
        print(f"Error parsing LLM response: {e}")
        return None

    except Exception as e:
        print(f"Unexpected error while parsing LLM response: {e}")
        return None

set_api_key(api_key, user=None)

Set the API key for the LLM provider.


api_key (str): The API key for the LLM provider.
user (Union[str, None]): The username

bool: True if the API key is successfully set, False otherwise.

ValueError: If the model name or correction model name is not set.
TypeError: If the LiteLLM object initialization fails.
Exception: If there is an unexpected error.
Source code in biochatter/llm_connect/llmlite.py
def set_api_key(self, api_key: str, user: str | None = None) -> bool:
    """Set the API key for the LLM provider.

    Args:
    ----
        api_key (str): The API key for the LLM provider.
        user (Union[str, None]): The username

    Returns:
    -------
        bool: True if the API key is successfully set, False otherwise.

    Raises:
    ------
        ValueError: If the model name or correction model name is not set.
        TypeError: If the LiteLLM object initialization fails.
        Exception: If there is an unexpected error.

    """
    try:
        if self.model_name is None:
            raise ValueError("Primary Model name is not set.")

        if self.ca_model_name is None:
            raise ValueError("Correction Model name is not set.")

        self.chat = self.get_litellm_object(api_key, self.model_name)
        if self.chat is None:
            raise TypeError("Failed to intialize primary agent chat object.")

        self.ca_chat = self.get_litellm_object(api_key, self.ca_model_name)
        if self.ca_chat is None:
            raise TypeError("Failed to intialize correcting agent chat object.")

        self.user = user
        if user == "community":
            self.usage_stats = get_stats(user=user)
        return True

    except (ValueError, TypeError):
        self.chat = None
        self.ca_chat = None
        return False
    except Exception:
        self.chat = None
        self.ca_chat = None
        return False

OllamaConversation

Bases: Conversation

Conversation class for the Ollama model.

Source code in biochatter/llm_connect/ollama.py
class OllamaConversation(Conversation):
    """Conversation class for the Ollama model."""

    def set_api_key(self, api_key: str, user: str | None = None) -> bool:
        """Set the API key for the Ollama API. Not implemented.

        Args:
        ----
            api_key (str): The API key for the Ollama API.

            user (str): The user for usage statistics.

        Returns:
        -------
            bool: True if the API key is valid, False otherwise.

        """
        err = "Ollama does not require an API key."
        raise NotImplementedError(err)

    def __init__(
        self,
        base_url: str,
        prompts: dict,
        model_name: str = "llama3",
        correct: bool = False,
        split_correction: bool = False,
    ) -> None:
        """Connect to an Ollama LLM via the Ollama/Langchain library.

        Set up a conversation with the user. Also initialise a second
        conversational agent to provide corrections to the model output, if
        necessary.

        Args:
        ----
            base_url (str): The base URL of the Ollama instance.

            prompts (dict): A dictionary of prompts to use for the conversation.

            model_name (str): The name of the model to use. Can be any model
                name available in your Ollama instance.

            correct (bool): Whether to correct the model output.

            split_correction (bool): Whether to correct the model output by
                splitting the output into sentences and correcting each sentence
                individually.

        """
        super().__init__(
            model_name=model_name,
            prompts=prompts,
            correct=correct,
            split_correction=split_correction,
        )
        self.model_name = model_name
        self.model = ChatOllama(
            base_url=base_url,
            model=self.model_name,
            temperature=0.0,
        )

        self.ca_model_name = "mixtral:latest"

        self.ca_model = ChatOllama(
            base_url=base_url,
            model_name=self.ca_model_name,
            temperature=0.0,
        )

    def append_system_message(self, message: str) -> None:
        """Override the system message addition.

        Ollama does not accept multiple system messages. Concatenate them if
        there are multiple.

        Args:
        ----
            message (str): The message to append.

        """
        # if there is not already a system message in self.messages
        if not any(isinstance(m, SystemMessage) for m in self.messages):
            self.messages.append(
                SystemMessage(
                    content=message,
                ),
            )
        else:
            # if there is a system message, append to the last one
            for i, msg in enumerate(self.messages):
                if isinstance(msg, SystemMessage):
                    self.messages[i].content += f"\n{message}"
                    break

    def append_ca_message(self, message: str) -> None:
        """Override the system message addition for the correcting agent.

        Ollama does not accept multiple system messages. Concatenate them if
        there are multiple.

        TODO this currently assumes that the correcting agent is the same model
        as the primary one.

        Args:
        ----
            message (str): The message to append.

        """
        # if there is not already a system message in self.messages
        if not any(isinstance(m, SystemMessage) for m in self.ca_messages):
            self.ca_messages.append(
                SystemMessage(
                    content=message,
                ),
            )
        else:
            # if there is a system message, append to the last one
            for i, msg in enumerate(self.ca_messages):
                if isinstance(msg, SystemMessage):
                    self.ca_messages[i].content += f"\n{message}"
                    break

    def _primary_query(self, **kwargs) -> tuple:
        """Query the Ollama client API with the user's message.

        Return the response using the message history (flattery system messages,
        prior conversation) as context. Correct the response if necessary.

        Returns
        -------
            tuple: A tuple containing the response from the Ollama API
            (formatted similarly to responses from the OpenAI API) and the token
            usage.

        """
        if kwargs:
            warnings.warn(f"Warning: {kwargs} are not used by this class", UserWarning)

        try:
            messages = self._create_history(self.messages)
            response = self.model.invoke(
                messages,
                # ,generate_config={"max_tokens": 2048, "temperature": 0},
            )
        except (
            openai._exceptions.APIError,
            openai._exceptions.OpenAIError,
            openai._exceptions.ConflictError,
            openai._exceptions.NotFoundError,
            openai._exceptions.APIStatusError,
            openai._exceptions.RateLimitError,
            openai._exceptions.APITimeoutError,
            openai._exceptions.BadRequestError,
            openai._exceptions.APIConnectionError,
            openai._exceptions.AuthenticationError,
            openai._exceptions.InternalServerError,
            openai._exceptions.PermissionDeniedError,
            openai._exceptions.UnprocessableEntityError,
            openai._exceptions.APIResponseValidationError,
        ) as e:
            return str(e), None
        response_dict = response.dict()
        msg = response_dict["content"]
        token_usage_raw = response_dict["response_metadata"]["eval_count"]
        token_usage = self._extract_total_tokens(token_usage_raw)

        self._update_usage_stats(self.model_name, token_usage_raw)

        self.append_ai_message(msg)

        return msg, token_usage

    def _create_history(self, messages: list) -> list:
        history = []
        for _, m in enumerate(messages):
            if isinstance(m, AIMessage):
                history.append(AIMessage(content=m.content))
            elif isinstance(m, HumanMessage):
                history.append(HumanMessage(content=m.content))
            elif isinstance(m, SystemMessage):
                history.append(SystemMessage(content=m.content))

        return history

    def _correct_response(self, msg: str) -> str:
        """Correct the response from the Ollama API.

        Send the response to a secondary language model. Optionally split the
        response into single sentences and correct each sentence individually.
        Update usage stats.

        Args:
        ----
            msg (str): The response from the model.

        Returns:
        -------
            str: The corrected response (or OK if no correction necessary).

        """
        ca_messages = self.ca_messages.copy()
        ca_messages.append(
            HumanMessage(
                content=msg,
            ),
        )
        ca_messages.append(
            SystemMessage(
                content="If there is nothing to correct, please respond with just 'OK', and nothing else!",
            ),
        )
        response = self.ca_model.invoke(
            chat_history=self._create_history(self.messages),
        ).dict()
        correction = response["content"]
        token_usage_raw = response["eval_count"]
        token_usage = self._extract_total_tokens(token_usage_raw)

        self._update_usage_stats(self.ca_model_name, token_usage_raw)

        return correction

    def _update_usage_stats(self, model: str, token_usage: dict) -> None:
        """Update redis database with token usage statistics.

        Use the usage_stats object with the increment method.

        Args:
        ----
            model (str): The model name.

            token_usage (dict): The token usage statistics.

        """

__init__(base_url, prompts, model_name='llama3', correct=False, split_correction=False)

Connect to an Ollama LLM via the Ollama/Langchain library.

Set up a conversation with the user. Also initialise a second conversational agent to provide corrections to the model output, if necessary.


base_url (str): The base URL of the Ollama instance.

prompts (dict): A dictionary of prompts to use for the conversation.

model_name (str): The name of the model to use. Can be any model
    name available in your Ollama instance.

correct (bool): Whether to correct the model output.

split_correction (bool): Whether to correct the model output by
    splitting the output into sentences and correcting each sentence
    individually.
Source code in biochatter/llm_connect/ollama.py
def __init__(
    self,
    base_url: str,
    prompts: dict,
    model_name: str = "llama3",
    correct: bool = False,
    split_correction: bool = False,
) -> None:
    """Connect to an Ollama LLM via the Ollama/Langchain library.

    Set up a conversation with the user. Also initialise a second
    conversational agent to provide corrections to the model output, if
    necessary.

    Args:
    ----
        base_url (str): The base URL of the Ollama instance.

        prompts (dict): A dictionary of prompts to use for the conversation.

        model_name (str): The name of the model to use. Can be any model
            name available in your Ollama instance.

        correct (bool): Whether to correct the model output.

        split_correction (bool): Whether to correct the model output by
            splitting the output into sentences and correcting each sentence
            individually.

    """
    super().__init__(
        model_name=model_name,
        prompts=prompts,
        correct=correct,
        split_correction=split_correction,
    )
    self.model_name = model_name
    self.model = ChatOllama(
        base_url=base_url,
        model=self.model_name,
        temperature=0.0,
    )

    self.ca_model_name = "mixtral:latest"

    self.ca_model = ChatOllama(
        base_url=base_url,
        model_name=self.ca_model_name,
        temperature=0.0,
    )

append_ca_message(message)

Override the system message addition for the correcting agent.

Ollama does not accept multiple system messages. Concatenate them if there are multiple.

TODO this currently assumes that the correcting agent is the same model as the primary one.


message (str): The message to append.
Source code in biochatter/llm_connect/ollama.py
def append_ca_message(self, message: str) -> None:
    """Override the system message addition for the correcting agent.

    Ollama does not accept multiple system messages. Concatenate them if
    there are multiple.

    TODO this currently assumes that the correcting agent is the same model
    as the primary one.

    Args:
    ----
        message (str): The message to append.

    """
    # if there is not already a system message in self.messages
    if not any(isinstance(m, SystemMessage) for m in self.ca_messages):
        self.ca_messages.append(
            SystemMessage(
                content=message,
            ),
        )
    else:
        # if there is a system message, append to the last one
        for i, msg in enumerate(self.ca_messages):
            if isinstance(msg, SystemMessage):
                self.ca_messages[i].content += f"\n{message}"
                break

append_system_message(message)

Override the system message addition.

Ollama does not accept multiple system messages. Concatenate them if there are multiple.


message (str): The message to append.
Source code in biochatter/llm_connect/ollama.py
def append_system_message(self, message: str) -> None:
    """Override the system message addition.

    Ollama does not accept multiple system messages. Concatenate them if
    there are multiple.

    Args:
    ----
        message (str): The message to append.

    """
    # if there is not already a system message in self.messages
    if not any(isinstance(m, SystemMessage) for m in self.messages):
        self.messages.append(
            SystemMessage(
                content=message,
            ),
        )
    else:
        # if there is a system message, append to the last one
        for i, msg in enumerate(self.messages):
            if isinstance(msg, SystemMessage):
                self.messages[i].content += f"\n{message}"
                break

set_api_key(api_key, user=None)

Set the API key for the Ollama API. Not implemented.


api_key (str): The API key for the Ollama API.

user (str): The user for usage statistics.

bool: True if the API key is valid, False otherwise.
Source code in biochatter/llm_connect/ollama.py
def set_api_key(self, api_key: str, user: str | None = None) -> bool:
    """Set the API key for the Ollama API. Not implemented.

    Args:
    ----
        api_key (str): The API key for the Ollama API.

        user (str): The user for usage statistics.

    Returns:
    -------
        bool: True if the API key is valid, False otherwise.

    """
    err = "Ollama does not require an API key."
    raise NotImplementedError(err)

OpenRouterConversation

Bases: LangChainConversation

Conversation class for the OpenRouter API.

Source code in biochatter/llm_connect/openrouter.py
class OpenRouterConversation(LangChainConversation):
    """Conversation class for the OpenRouter API."""

    def __init__(self, model_name: str, prompts: dict, **kwargs):
        super().__init__(model_name, "", prompts, **kwargs)

    def set_api_key(self, api_key: str | None = None, user: str | None = None) -> bool:
        """Set the API key for the model provider.

        If the key is valid, initialise the conversational agent. Optionally set
        the user for usage statistics.

        Args:
        ----
            api_key (str): The API key for the model provider.

            user (str, optional): The user for usage statistics. If provided and
                equals "community", will track usage stats.

        Returns:
        -------
            bool: True if the API key is valid, False otherwise.

        """
        self.user = user

        try:
            self.chat = ChatOpenRouter(
                model_name=self.model_name,
                temperature=0,
            )
            self.ca_chat = ChatOpenRouter(
                model_name=self.model_name,
                temperature=0,
            )

            # if binding happens here, tools will be available for all messages
            if self.tools:
                self.bind_tools(self.tools)

            return True

        except Exception:
            self._chat = None
            self._ca_chat = None
            return False

set_api_key(api_key=None, user=None)

Set the API key for the model provider.

If the key is valid, initialise the conversational agent. Optionally set the user for usage statistics.


api_key (str): The API key for the model provider.

user (str, optional): The user for usage statistics. If provided and
    equals "community", will track usage stats.

bool: True if the API key is valid, False otherwise.
Source code in biochatter/llm_connect/openrouter.py
def set_api_key(self, api_key: str | None = None, user: str | None = None) -> bool:
    """Set the API key for the model provider.

    If the key is valid, initialise the conversational agent. Optionally set
    the user for usage statistics.

    Args:
    ----
        api_key (str): The API key for the model provider.

        user (str, optional): The user for usage statistics. If provided and
            equals "community", will track usage stats.

    Returns:
    -------
        bool: True if the API key is valid, False otherwise.

    """
    self.user = user

    try:
        self.chat = ChatOpenRouter(
            model_name=self.model_name,
            temperature=0,
        )
        self.ca_chat = ChatOpenRouter(
            model_name=self.model_name,
            temperature=0,
        )

        # if binding happens here, tools will be available for all messages
        if self.tools:
            self.bind_tools(self.tools)

        return True

    except Exception:
        self._chat = None
        self._ca_chat = None
        return False

WasmConversation

Bases: Conversation

Conversation class for the wasm model.

Source code in biochatter/llm_connect/misc.py
class WasmConversation(Conversation):
    """Conversation class for the wasm model."""

    def __init__(
        self,
        model_name: str,
        prompts: dict,
        correct: bool = False,
        split_correction: bool = False,
    ) -> None:
        """Initialize the WasmConversation class.

        This class is used to return the complete query as a string to be used
        in the frontend running the wasm model. It does not call the API itself,
        but updates the message history similarly to the other conversation
        classes. It overrides the `query` method from the `Conversation` class
        to return a plain string that contains the entire message for the model
        as the first element of the tuple. The second and third elements are
        `None` as there is no token usage or correction for the wasm model.

        """
        super().__init__(
            model_name=model_name,
            prompts=prompts,
            correct=correct,
            split_correction=split_correction,
        )

    def query(self, text: str) -> tuple:
        """Return the entire message history as a single string.

        This is the message that is sent to the wasm model.

        Args:
        ----
            text (str): The user query.

        Returns:
        -------
            tuple: A tuple containing the message history as a single string,
                and `None` for the second and third elements of the tuple.

        """
        self.append_user_message(text)

        self._inject_context(text)

        return (self._primary_query(), None, None)

    def _primary_query(self):
        """Concatenate all messages in the conversation.

        Build a single string from all messages in the conversation.
        Currently discards information about roles (system, user).

        Returns
        -------
            str: A single string from all messages in the conversation.

        """
        return "\n".join([m.content for m in self.messages])

    def _correct_response(self, msg: str) -> str:
        """Do not use for the wasm model."""
        return "ok"

    def set_api_key(self, api_key: str, user: str | None = None) -> bool:
        """Do not use for the wasm model."""
        return True

__init__(model_name, prompts, correct=False, split_correction=False)

Initialize the WasmConversation class.

This class is used to return the complete query as a string to be used in the frontend running the wasm model. It does not call the API itself, but updates the message history similarly to the other conversation classes. It overrides the query method from the Conversation class to return a plain string that contains the entire message for the model as the first element of the tuple. The second and third elements are None as there is no token usage or correction for the wasm model.

Source code in biochatter/llm_connect/misc.py
def __init__(
    self,
    model_name: str,
    prompts: dict,
    correct: bool = False,
    split_correction: bool = False,
) -> None:
    """Initialize the WasmConversation class.

    This class is used to return the complete query as a string to be used
    in the frontend running the wasm model. It does not call the API itself,
    but updates the message history similarly to the other conversation
    classes. It overrides the `query` method from the `Conversation` class
    to return a plain string that contains the entire message for the model
    as the first element of the tuple. The second and third elements are
    `None` as there is no token usage or correction for the wasm model.

    """
    super().__init__(
        model_name=model_name,
        prompts=prompts,
        correct=correct,
        split_correction=split_correction,
    )

query(text)

Return the entire message history as a single string.

This is the message that is sent to the wasm model.


text (str): The user query.

tuple: A tuple containing the message history as a single string,
    and `None` for the second and third elements of the tuple.
Source code in biochatter/llm_connect/misc.py
def query(self, text: str) -> tuple:
    """Return the entire message history as a single string.

    This is the message that is sent to the wasm model.

    Args:
    ----
        text (str): The user query.

    Returns:
    -------
        tuple: A tuple containing the message history as a single string,
            and `None` for the second and third elements of the tuple.

    """
    self.append_user_message(text)

    self._inject_context(text)

    return (self._primary_query(), None, None)

set_api_key(api_key, user=None)

Do not use for the wasm model.

Source code in biochatter/llm_connect/misc.py
def set_api_key(self, api_key: str, user: str | None = None) -> bool:
    """Do not use for the wasm model."""
    return True

XinferenceConversation

Bases: Conversation

Conversation class for the Xinference deployment.

Source code in biochatter/llm_connect/xinference.py
class XinferenceConversation(Conversation):
    """Conversation class for the Xinference deployment."""

    def __init__(
        self,
        base_url: str,
        prompts: dict,
        model_name: str = "auto",
        correct: bool = False,
        split_correction: bool = False,
    ) -> None:
        """Connect to an open-source LLM via the Xinference client.

        Connect to a running Xinference deployment and set up a conversation
        with the user. Also initialise a second conversational agent to
        provide corrections to the model output, if necessary.

        Args:
        ----
            base_url (str): The base URL of the Xinference instance (should not
            include the /v1 part).

            prompts (dict): A dictionary of prompts to use for the conversation.

            model_name (str): The name of the model to use. Will be mapped to
            the according uid from the list of available models. Can be set to
            "auto" to use the first available model.

            correct (bool): Whether to correct the model output.

            split_correction (bool): Whether to correct the model output by
            splitting the output into sentences and correcting each sentence
            individually.

        """
        # Shaohong: Please keep this xinference importing code here, so that,
        # we don't need to depend on xinference if we dont need it (xinference
        # is expensive to install)
        from xinference.client import Client

        super().__init__(
            model_name=model_name,
            prompts=prompts,
            correct=correct,
            split_correction=split_correction,
        )
        self.client = Client(base_url=base_url)

        self.models = {}
        self.load_models()

        self.ca_model_name = model_name

        self.set_api_key()

        # TODO make accessible by drop-down

    def load_models(self) -> None:
        """Load the models from the Xinference client."""
        for id, model in self.client.list_models().items():
            model["id"] = id
            self.models[model["model_name"]] = model

    def append_system_message(self, message: str) -> None:
        """Override the system message addition.

        Xinference does not accept multiple system messages. We concatenate them
        if there are multiple.

        Args:
        ----
            message (str): The message to append.

        """
        # if there is not already a system message in self.messages
        if not any(isinstance(m, SystemMessage) for m in self.messages):
            self.messages.append(
                SystemMessage(
                    content=message,
                ),
            )
        else:
            # if there is a system message, append to the last one
            for i, msg in enumerate(self.messages):
                if isinstance(msg, SystemMessage):
                    self.messages[i].content += f"\n{message}"
                    break

    def append_ca_message(self, message: str) -> None:
        """Override the system message addition for the correcting agent.

        Xinference does not accept multiple system messages. We concatenate them
        if there are multiple.

        TODO this currently assumes that the correcting agent is the same model
        as the primary one.

        Args:
        ----
            message (str): The message to append.

        """
        # if there is not already a system message in self.messages
        if not any(isinstance(m, SystemMessage) for m in self.ca_messages):
            self.ca_messages.append(
                SystemMessage(
                    content=message,
                ),
            )
        else:
            # if there is a system message, append to the last one
            for i, msg in enumerate(self.ca_messages):
                if isinstance(msg, SystemMessage):
                    self.ca_messages[i].content += f"\n{message}"
                    break

    def _primary_query(self, **kwargs) -> tuple:
        """Query the Xinference client API.

        Use the user's message and return the response using the message history
        (flattery system messages, prior conversation) as context. Correct the
        response if necessary.

        LLaMA2 architecture does not accept separate system messages, so we
        concatenate the system message with the user message to form the prompt.
        'LLaMA enforces a strict rule that chats should alternate
        user/assistant/user/assistant, and the system message, if present,
        should be embedded into the first user message.' (from
        https://discuss.huggingface.co/t/issue-with-llama-2-chat-template-and-out-of-date-documentation/61645/3)

        Returns
        -------
            tuple: A tuple containing the response from the Xinference API
            (formatted similarly to responses from the OpenAI API) and the token
            usage.

        """
        if kwargs:
            warnings.warn(f"Warning: {kwargs} are not used by this class", UserWarning)

        try:
            history = self._create_history()
            # TODO this is for LLaMA2 arch, may be different for newer models
            prompt = history.pop()
            response = self.model.chat(
                prompt=prompt["content"],
                chat_history=history,
                generate_config={"max_tokens": 2048, "temperature": 0},
            )
        except (
            openai._exceptions.APIError,
            openai._exceptions.OpenAIError,
            openai._exceptions.ConflictError,
            openai._exceptions.NotFoundError,
            openai._exceptions.APIStatusError,
            openai._exceptions.RateLimitError,
            openai._exceptions.APITimeoutError,
            openai._exceptions.BadRequestError,
            openai._exceptions.APIConnectionError,
            openai._exceptions.AuthenticationError,
            openai._exceptions.InternalServerError,
            openai._exceptions.PermissionDeniedError,
            openai._exceptions.UnprocessableEntityError,
            openai._exceptions.APIResponseValidationError,
        ) as e:
            return str(e), None

        msg = response["choices"][0]["message"]["content"]
        token_usage_raw = response["usage"]
        token_usage = self._extract_total_tokens(token_usage_raw)

        self._update_usage_stats(self.model_name, token_usage_raw)

        self.append_ai_message(msg)

        return msg, token_usage

    def _create_history(self) -> list:
        """Create a history of messages from the conversation.

        Returns
        -------
            list: A list of messages from the conversation.

        """
        history = []
        # extract text components from message contents
        msg_texts = [m.content[0]["text"] if isinstance(m.content, list) else m.content for m in self.messages]

        # check if last message is an image message
        is_image_message = False
        if isinstance(self.messages[-1].content, list):
            is_image_message = self.messages[-1].content[1]["type"] == "image_url"

        # find location of last AI message (if any)
        last_ai_message = None
        for i, m in enumerate(self.messages):
            if isinstance(m, AIMessage):
                last_ai_message = i

        # concatenate all messages before the last AI message into one message
        if last_ai_message:
            history.append(
                {
                    "role": "user",
                    "content": "\n".join(
                        [m for m in msg_texts[:last_ai_message]],
                    ),
                },
            )
            # then append the last AI message
            history.append(
                {
                    "role": "assistant",
                    "content": msg_texts[last_ai_message],
                },
            )

            # then concatenate all messages after that
            # into one HumanMessage
            history.append(
                {
                    "role": "user",
                    "content": "\n".join(
                        [m for m in msg_texts[last_ai_message + 1 :]],
                    ),
                },
            )

        # if there is no AI message, concatenate all messages into one user
        # message
        else:
            history.append(
                {
                    "role": "user",
                    "content": "\n".join([m for m in msg_texts[:]]),
                },
            )

        # if the last message is an image message, add the image to the history
        if is_image_message:
            history[-1].content = [
                {"type": "text", "text": history[-1].content},
                {
                    "type": "image_url",
                    "image_url": {
                        "url": self.messages[-1].content[1]["image_url"]["url"],
                    },
                },
            ]
        return history

    def _correct_response(self, msg: str) -> str:
        """Correct the response from the Xinference API.

        Send the response to a secondary language model. Optionally split the
        response into single sentences and correct each sentence individually.
        Update usage stats.

        Args:
        ----
            msg (str): The response from the model.

        Returns:
        -------
            str: The corrected response (or OK if no correction necessary).

        """
        ca_messages = self.ca_messages.copy()
        ca_messages.append(
            HumanMessage(
                content=msg,
            ),
        )
        ca_messages.append(
            SystemMessage(
                content="If there is nothing to correct, please respond with just 'OK', and nothing else!",
            ),
        )
        history = []
        for m in self.messages:
            if isinstance(m, SystemMessage):
                history.append({"role": "system", "content": m.content})
            elif isinstance(m, HumanMessage):
                history.append({"role": "user", "content": m.content})
            elif isinstance(m, AIMessage):
                history.append({"role": "assistant", "content": m.content})
        prompt = history.pop()
        response = self.ca_model.chat(
            prompt=prompt["content"],
            chat_history=history,
            generate_config={"max_tokens": 2048, "temperature": 0},
        )

        correction = response["choices"][0]["message"]["content"]
        token_usage_raw = response["usage"]
        token_usage = self._extract_total_tokens(token_usage_raw)

        self._update_usage_stats(self.ca_model_name, token_usage_raw)

        return correction

    def _update_usage_stats(self, model: str, token_usage: dict) -> None:
        """Update redis database with token usage statistics.

        Use the usage_stats object with the increment method.

        Args:
        ----
            model (str): The model name.

            token_usage (dict): The token usage statistics.

        """

    def set_api_key(self) -> bool:
        """Try to get the Xinference model from the client API.

        If the model is found, initialise the conversational agent. If the model
        is not found, `get_model` will raise a RuntimeError.

        Returns
        -------
            bool: True if the model is found, False otherwise.

        """
        try:
            if self.model_name is None or self.model_name == "auto":
                self.model_name = self.list_models_by_type("chat")[0]
            self.model = self.client.get_model(
                self.models[self.model_name]["id"],
            )

            if self.ca_model_name is None or self.ca_model_name == "auto":
                self.ca_model_name = self.list_models_by_type("chat")[0]
            self.ca_model = self.client.get_model(
                self.models[self.ca_model_name]["id"],
            )
            return True

        except RuntimeError:
            self._chat = None
            self._ca_chat = None
            return False

    def list_models_by_type(self, model_type: str) -> list[str]:
        """List the models by type.

        Args:
        ----
            model_type (str): The type of model to list.

        Returns:
        -------
            list[str]: A list of model names.

        """
        names = []
        if model_type in ["embed", "embedding"]:
            for name, model in self.models.items():
                if "model_ability" in model:
                    if "embed" in model["model_ability"]:
                        names.append(name)
                elif model["model_type"] == "embedding":
                    names.append(name)
            return names
        for name, model in self.models.items():
            if "model_ability" in model:
                if model_type in model["model_ability"]:
                    names.append(name)
            elif model["model_type"] == model_type:
                names.append(name)
        return names

__init__(base_url, prompts, model_name='auto', correct=False, split_correction=False)

Connect to an open-source LLM via the Xinference client.

Connect to a running Xinference deployment and set up a conversation with the user. Also initialise a second conversational agent to provide corrections to the model output, if necessary.


base_url (str): The base URL of the Xinference instance (should not
include the /v1 part).

prompts (dict): A dictionary of prompts to use for the conversation.

model_name (str): The name of the model to use. Will be mapped to
the according uid from the list of available models. Can be set to
"auto" to use the first available model.

correct (bool): Whether to correct the model output.

split_correction (bool): Whether to correct the model output by
splitting the output into sentences and correcting each sentence
individually.
Source code in biochatter/llm_connect/xinference.py
def __init__(
    self,
    base_url: str,
    prompts: dict,
    model_name: str = "auto",
    correct: bool = False,
    split_correction: bool = False,
) -> None:
    """Connect to an open-source LLM via the Xinference client.

    Connect to a running Xinference deployment and set up a conversation
    with the user. Also initialise a second conversational agent to
    provide corrections to the model output, if necessary.

    Args:
    ----
        base_url (str): The base URL of the Xinference instance (should not
        include the /v1 part).

        prompts (dict): A dictionary of prompts to use for the conversation.

        model_name (str): The name of the model to use. Will be mapped to
        the according uid from the list of available models. Can be set to
        "auto" to use the first available model.

        correct (bool): Whether to correct the model output.

        split_correction (bool): Whether to correct the model output by
        splitting the output into sentences and correcting each sentence
        individually.

    """
    # Shaohong: Please keep this xinference importing code here, so that,
    # we don't need to depend on xinference if we dont need it (xinference
    # is expensive to install)
    from xinference.client import Client

    super().__init__(
        model_name=model_name,
        prompts=prompts,
        correct=correct,
        split_correction=split_correction,
    )
    self.client = Client(base_url=base_url)

    self.models = {}
    self.load_models()

    self.ca_model_name = model_name

    self.set_api_key()

append_ca_message(message)

Override the system message addition for the correcting agent.

Xinference does not accept multiple system messages. We concatenate them if there are multiple.

TODO this currently assumes that the correcting agent is the same model as the primary one.


message (str): The message to append.
Source code in biochatter/llm_connect/xinference.py
def append_ca_message(self, message: str) -> None:
    """Override the system message addition for the correcting agent.

    Xinference does not accept multiple system messages. We concatenate them
    if there are multiple.

    TODO this currently assumes that the correcting agent is the same model
    as the primary one.

    Args:
    ----
        message (str): The message to append.

    """
    # if there is not already a system message in self.messages
    if not any(isinstance(m, SystemMessage) for m in self.ca_messages):
        self.ca_messages.append(
            SystemMessage(
                content=message,
            ),
        )
    else:
        # if there is a system message, append to the last one
        for i, msg in enumerate(self.ca_messages):
            if isinstance(msg, SystemMessage):
                self.ca_messages[i].content += f"\n{message}"
                break

append_system_message(message)

Override the system message addition.

Xinference does not accept multiple system messages. We concatenate them if there are multiple.


message (str): The message to append.
Source code in biochatter/llm_connect/xinference.py
def append_system_message(self, message: str) -> None:
    """Override the system message addition.

    Xinference does not accept multiple system messages. We concatenate them
    if there are multiple.

    Args:
    ----
        message (str): The message to append.

    """
    # if there is not already a system message in self.messages
    if not any(isinstance(m, SystemMessage) for m in self.messages):
        self.messages.append(
            SystemMessage(
                content=message,
            ),
        )
    else:
        # if there is a system message, append to the last one
        for i, msg in enumerate(self.messages):
            if isinstance(msg, SystemMessage):
                self.messages[i].content += f"\n{message}"
                break

list_models_by_type(model_type)

List the models by type.


model_type (str): The type of model to list.

list[str]: A list of model names.
Source code in biochatter/llm_connect/xinference.py
def list_models_by_type(self, model_type: str) -> list[str]:
    """List the models by type.

    Args:
    ----
        model_type (str): The type of model to list.

    Returns:
    -------
        list[str]: A list of model names.

    """
    names = []
    if model_type in ["embed", "embedding"]:
        for name, model in self.models.items():
            if "model_ability" in model:
                if "embed" in model["model_ability"]:
                    names.append(name)
            elif model["model_type"] == "embedding":
                names.append(name)
        return names
    for name, model in self.models.items():
        if "model_ability" in model:
            if model_type in model["model_ability"]:
                names.append(name)
        elif model["model_type"] == model_type:
            names.append(name)
    return names

load_models()

Load the models from the Xinference client.

Source code in biochatter/llm_connect/xinference.py
def load_models(self) -> None:
    """Load the models from the Xinference client."""
    for id, model in self.client.list_models().items():
        model["id"] = id
        self.models[model["model_name"]] = model

set_api_key()

Try to get the Xinference model from the client API.

If the model is found, initialise the conversational agent. If the model is not found, get_model will raise a RuntimeError.

Returns
bool: True if the model is found, False otherwise.
Source code in biochatter/llm_connect/xinference.py
def set_api_key(self) -> bool:
    """Try to get the Xinference model from the client API.

    If the model is found, initialise the conversational agent. If the model
    is not found, `get_model` will raise a RuntimeError.

    Returns
    -------
        bool: True if the model is found, False otherwise.

    """
    try:
        if self.model_name is None or self.model_name == "auto":
            self.model_name = self.list_models_by_type("chat")[0]
        self.model = self.client.get_model(
            self.models[self.model_name]["id"],
        )

        if self.ca_model_name is None or self.ca_model_name == "auto":
            self.ca_model_name = self.list_models_by_type("chat")[0]
        self.ca_model = self.client.get_model(
            self.models[self.ca_model_name]["id"],
        )
        return True

    except RuntimeError:
        self._chat = None
        self._ca_chat = None
        return False