Vectorstore agent module

Here we handle connections and management of various vectorstore services and semantic seach.

VectorDatabaseAgentMilvus

The VectorDatabaseAgentMilvus class manages vector databases in a connected host database. It manages an embedding collection _col_embeddings:langchain.vectorstores.Milvus, which is the main information on the embedded text fragments and the basis for similarity search, and a metadata collection _col_metadata:pymilvus.Collection, which stores the metadata of the embedded text fragments. A typical workflow includes the following operations:

1. connect to a host using connect()
2. get all documents in the active database using get_all_documents()
3. save a number of fragments, usually from a specific document, using store_embeddings()
4. do similarity search among all fragments of the currently active database using similarity_search()
5. remove a document from the currently active database using remove_document()

Source code in biochatter/vectorstore_agent.py

class VectorDatabaseAgentMilvus:
    """
    The VectorDatabaseAgentMilvus class manages vector databases in a connected
    host database. It manages an embedding collection
    `_col_embeddings:langchain.vectorstores.Milvus`, which is the main
    information on the embedded text fragments and the basis for similarity
    search, and a metadata collection `_col_metadata:pymilvus.Collection`, which
    stores the metadata of the embedded text fragments. A typical workflow
    includes the following operations:

    1. connect to a host using `connect()`
    2. get all documents in the active database using `get_all_documents()`
    3. save a number of fragments, usually from a specific document, using
        `store_embeddings()`
    4. do similarity search among all fragments of the currently active database
        using `similarity_search()`
    5. remove a document from the currently active database using
        `remove_document()`
    """

    def __init__(
        self,
        embedding_func: OpenAIEmbeddings,
        connection_args: Optional[dict] = None,
        embedding_collection_name: Optional[str] = None,
        metadata_collection_name: Optional[str] = None,
    ):
        """
        Args:
            embedding_func OpenAIEmbeddings: Function used to embed the text

            connection_args Optional dict: args to connect Vector Database

            embedding_collection_name Optional str: exposed for test

            metadata_collection_name Optional str: exposed for test
        """
        self._embedding_func = embedding_func
        self._col_embeddings: Optional[Milvus] = None
        self._col_metadata: Optional[Collection] = None
        self._connection_args = validate_connection_args(connection_args)
        self._embedding_name = (
            embedding_collection_name or DOCUMENT_EMBEDDINGS_COLLECTION_NAME
        )
        self._metadata_name = (
            metadata_collection_name or DOCUMENT_METADATA_COLLECTION_NAME
        )

    def connect(self) -> None:
        """
        Connect to a host and read two document collections (the default names
        are `DocumentEmbeddings` and `DocumentMetadata`) in the currently active
        database (default database name is `default`); if those document
        collections don't exist, create the two collections.
        """
        self._connect(**self._connection_args)
        self._init_host()

    def _connect(self, host: str, port: str, user: str, password: str) -> None:
        self.alias = self._create_connection_alias(host, port, user, password)

    def _init_host(self) -> None:
        """
        Initialize host. Will read/create document collection inside currently
        active database.
        """
        self._create_collections()

    def _create_connection_alias(
        self, host: str, port: str, user: str, password: str
    ) -> str:
        """
        Connect to host and create a connection alias for metadata collection
        using a random uuid.

        Args:
            host (str): host ip address
            port (str): host port

        Returns:
            str: connection alias
        """
        alias = uuid.uuid4().hex
        try:
            connections.connect(
                host=host, port=port, user=user, password=password, alias=alias
            )
            logger.debug(f"Created new connection using: {alias}")
            return alias
        except MilvusException as e:
            logger.error(f"Failed to create  new connection using: {alias}")
            raise e

    def _create_collections(self) -> None:
        """
        Create or load the embedding and metadata collections from the currently
        active database.
        """
        embedding_exists = utility.has_collection(
            self._embedding_name, using=self.alias
        )
        meta_exists = utility.has_collection(
            self._metadata_name,
            using=self.alias,
        )

        if embedding_exists:
            self._load_embeddings_collection()
        else:
            self._create_embeddings_collection()

        if meta_exists:
            self._load_metadata_collection()
        else:
            self._create_metadata_collection()

        self._create_metadata_collection_index()
        self._col_metadata.load()

    def _load_embeddings_collection(self) -> None:
        """
        Load embeddings collection from currently active database.
        """
        try:
            self._col_embeddings = Milvus(
                embedding_function=self._embedding_func,
                collection_name=self._embedding_name,
                connection_args=self._connection_args,
            )
        except MilvusException as e:
            logger.error(
                f"Failed to load embeddings collection {self._embedding_name}."
            )
            raise e

    def _create_embeddings_collection(self) -> None:
        """
        Create embedding collection.
        All fields: "meta_id", "vector"
        """
        try:
            self._col_embeddings = Milvus(
                embedding_function=self._embedding_func,
                collection_name=self._embedding_name,
                connection_args=self._connection_args,
            )
        except MilvusException as e:
            logger.error(
                f"Failed to create embeddings collection {self._embedding_name}"
            )
            raise e

    def _load_metadata_collection(self) -> None:
        """
        Load metadata collection from currently active database.
        """
        self._col_metadata = Collection(
            self._metadata_name,
            using=self.alias,
        )
        self._col_metadata.load()

    def _create_metadata_collection(self) -> None:
        """
        Create metadata collection.

        All fields: "id", "name", "author", "title", "format", "subject",
        "creator", "producer", "creationDate", "modDate", "source", "embedding",
        "isDeleted".

        As the vector database requires a vector field, we will create a fake
        vector "embedding". The field "isDeleted" is used to specify if the
        document is deleted.
        """
        doc_id = FieldSchema(
            name="id", dtype=DataType.INT64, is_primary=True, auto_id=True
        )
        doc_name = FieldSchema(
            name="name", dtype=DataType.VARCHAR, max_length=255
        )
        doc_author = FieldSchema(
            name="author", dtype=DataType.VARCHAR, max_length=255
        )
        doc_title = FieldSchema(
            name="title", dtype=DataType.VARCHAR, max_length=1000
        )
        doc_format = FieldSchema(
            name="format", dtype=DataType.VARCHAR, max_length=255
        )
        doc_subject = FieldSchema(
            name="subject", dtype=DataType.VARCHAR, max_length=255
        )
        doc_creator = FieldSchema(
            name="creator", dtype=DataType.VARCHAR, max_length=255
        )
        doc_producer = FieldSchema(
            name="producer", dtype=DataType.VARCHAR, max_length=255
        )
        doc_creationDate = FieldSchema(
            name="creationDate", dtype=DataType.VARCHAR, max_length=64
        )
        doc_modDate = FieldSchema(
            name="modDate", dtype=DataType.VARCHAR, max_length=64
        )
        doc_source = FieldSchema(
            name="source", dtype=DataType.VARCHAR, max_length=1000
        )
        embedding = FieldSchema(
            name="embedding",
            dtype=DataType.FLOAT_VECTOR,
            dim=METADATA_VECTOR_DIM,
        )
        isDeleted = FieldSchema(
            name="isDeleted",
            dtype=DataType.BOOL,
        )
        fields = [
            doc_id,
            doc_name,
            doc_author,
            doc_title,
            doc_format,
            doc_subject,
            doc_creator,
            doc_producer,
            doc_creationDate,
            doc_modDate,
            doc_source,
            embedding,
            isDeleted,
        ]
        schema = CollectionSchema(fields=fields)
        try:
            self._col_metadata = Collection(
                name=self._metadata_name, schema=schema, using=self.alias
            )
        except MilvusException as e:
            logger.error(f"Failed to create collection {self._metadata_name}")
            raise e

    def _create_metadata_collection_index(self) -> None:
        """
        Create index for metadata collection in currently active database.
        """
        if (
            not isinstance(self._col_metadata, Collection)
            or len(self._col_metadata.indexes) > 0
        ):
            return

        index_params = {
            "metric_type": "L2",
            "index_type": "HNSW",
            "params": {"M": 8, "efConstruction": 64},
        }

        try:
            self._col_metadata.create_index(
                field_name="embedding",
                index_params=index_params,
                using=self.alias,
            )
        except MilvusException as e:
            logger.error(
                "Failed to create index for meta collection "
                f"{self._metadata_name}."
            )
            raise e

    def _insert_data(self, documents: list[Document]) -> str:
        """
        Insert documents into the currently active database.

        Args:
            documents (List[Documents]): documents array, usually from
                DocumentReader.load_document, DocumentReader.document_from_pdf,
                DocumentReader.document_from_txt

        Returns:
            str: document id
        """
        if len(documents) == 0:
            return None
        metadata = [documents[0].metadata]
        aligned_metadata = align_metadata(metadata)
        try:
            result = self._col_metadata.insert(aligned_metadata)
            meta_id = str(result.primary_keys[0])
            self._col_metadata.flush()
        except MilvusException as e:
            logger.error(f"Failed to insert meta data")
            raise e
        aligned_docs = align_embeddings(documents, meta_id)
        try:
            # As we passed collection_name, documents will be added to existed collection
            self._col_embeddings = Milvus.from_documents(
                embedding=self._embedding_func,
                collection_name=self._embedding_name,
                connection_args=self._connection_args,
                documents=aligned_docs,
            )
        except MilvusException as e:
            logger.error(
                "Failed to insert data to embedding collection "
                f"{self._embedding_name}."
            )
            raise e
        return meta_id

    def store_embeddings(self, documents: list[Document]) -> str:
        """
        Store documents in the currently active database.

        Args:
            documents (List[Documents]): documents array, usually from
                DocumentReader.load_document, DocumentReader.document_from_pdf,
                DocumentReader.document_from_txt

        Returns:
            str: document id
        """
        if len(documents) == 0:
            return
        return self._insert_data(documents)

    def _build_embedding_search_expression(
        self, meta_ids: list[dict]
    ) -> Optional[str]:
        """
        Build search expression for embedding collection. The generated
        expression follows the pattern: "meta_id in [{id1}, {id2}, ...]

        Args:
            meta_ids: the array of metadata id in metadata collection

        Returns:
            str: search expression or None
        """
        if len(meta_ids) == 0:
            return "meta_id in []"
        built_expr = """meta_id in ["""
        for item in meta_ids:
            id = f'"{item["id"]}",'
            built_expr += id
        built_expr = built_expr[:-1]
        built_expr += """]"""
        return built_expr

    def _join_embedding_and_metadata_results(
        self, result_embedding: list[Document], result_meta: list[dict]
    ) -> list[Document]:
        """
        Join the search results of embedding collection and results of metadata.

        Args:
            result_embedding (List[Document]): search result of embedding
                collection

            result_meta (List[Dict]): search result of metadata collection

        Returns:
            List[Document]: combined results like
                [{page_content: str, metadata: {...}}]
        """

        def _find_metadata_by_id(
            metadata: list[dict], id: str
        ) -> Optional[dict]:
            for d in metadata:
                if str(d["id"]) == id:
                    return d
            return None

        joined_docs = []
        for res in result_embedding:
            found = _find_metadata_by_id(result_meta, res.metadata["meta_id"])
            if found is None:  # discard
                logger.error(
                    f"Failed to join meta_id {res.metadata['meta_id']}"
                )
                continue
            joined_docs.append(
                Document(page_content=res.page_content, metadata=found)
            )
        return joined_docs

    @staticmethod
    def _build_meta_col_query_expr_for_all_documents(
        doc_ids: Optional[list[str]] = None,
    ) -> str:
        """
        Build metadata collection query expression to obtain all documents.

        Args:
            doc_ids: the list of document ids (metadata ids), if thie argument is None,
                     that is, the query is to get all undeleted documents in metadata collection.
                     Otherwise, the query is to getr all undeleted documents form provided doc_ids

        Returns:
            query: str
        """
        expr = (
            f"id in {doc_ids} and isDeleted == false"
            if doc_ids is not None
            else "isDeleted == false"
        )
        return expr.replace('"', "").replace("'", "")

    def similarity_search(
        self, query: str, k: int = 3, doc_ids: list[str] = None
    ) -> list[Document]:
        """
        Perform similarity search insider the currently active database
        according to the input query.

        This method will:
        1. get all non-deleted meta_id and build to search expression for
            the currently active embedding collection
        2. do similarity search in the embedding collection
        3. combine metadata and embeddings

        Args:
            query (str): query string

            k (int): the number of results to return

            doc_ids(List[str] optional): the list of document ids, do similarity search across the
            specified documents

        Returns:
            List[Document]: search results
        """
        result_metadata = []
        expr = VectorDatabaseAgentMilvus._build_meta_col_query_expr_for_all_documents(
            doc_ids
        )
        result_metadata = self._col_metadata.query(
            expr=expr, output_fields=METADATA_FIELDS
        )
        expr = self._build_embedding_search_expression(result_metadata)
        result_embedding = self._col_embeddings.similarity_search(
            query=query, k=k, expr=expr
        )
        return self._join_embedding_and_metadata_results(
            result_embedding, result_metadata
        )

    def remove_document(
        self, doc_id: str, doc_ids: Optional[list[str]] = None
    ) -> bool:
        """
        Remove the document include meta data and its embeddings.

        Args:
            doc_id (str): the document to be deleted

            doc_ids(List[str] optional): the list of document ids, defines documents scope
            within which remove operation occurs.

        Returns:
            bool: True if the document is deleted, False otherwise
        """
        if not self._col_metadata:
            return False
        if doc_ids is not None and (
            len(doc_ids) == 0 or (len(doc_ids) > 0 and not doc_id in doc_ids)
        ):
            return False
        try:
            expr = f"id in [{doc_id}]"
            res = self._col_metadata.query(
                expr=expr, output_fields=METADATA_FIELDS
            )
            if len(res) == 0:
                return False
            del_res = self._col_metadata.delete(expr)
            self._col_metadata.flush()

            res = self._col_embeddings.col.query(f'meta_id in ["{doc_id}"]')
            if len(res) == 0:
                return True
            ids = [item["pk"] for item in res]
            embedding_expr = f"pk in {ids}"
            del_res = self._col_embeddings.col.delete(expr=embedding_expr)
            self._col_embeddings.col.flush()
            return True
        except MilvusException as e:
            logger.error(e)
            raise e

    def get_all_documents(
        self, doc_ids: Optional[list[str]] = None
    ) -> list[dict]:
        """
        Get all non-deleted documents from the currently active database.

        Args:
            doc_ids(List[str] optional): the list of document ids, defines documents scope within
            which the operation of obaining all documents occurs

        Returns:
            List[Dict]: the metadata of all non-deleted documents in the form
                [{{id}, {author}, {source}, ...}]
        """
        try:
            expr = VectorDatabaseAgentMilvus._build_meta_col_query_expr_for_all_documents(
                doc_ids
            )
            result_metadata = self._col_metadata.query(
                expr=expr, output_fields=METADATA_FIELDS
            )
            return result_metadata
        except MilvusException as e:
            logger.error(e)
            raise e

__init__(embedding_func, connection_args=None, embedding_collection_name=None, metadata_collection_name=None)

Parameters:

Name	Type	Description	Default
embedding_func	OpenAIEmbeddings	Function used to embed the text	required
connection_args	Optional dict	args to connect Vector Database	None
embedding_collection_name	Optional str	exposed for test	None
metadata_collection_name	Optional str	exposed for test	None

Source code in biochatter/vectorstore_agent.py

def __init__(
    self,
    embedding_func: OpenAIEmbeddings,
    connection_args: Optional[dict] = None,
    embedding_collection_name: Optional[str] = None,
    metadata_collection_name: Optional[str] = None,
):
    """
    Args:
        embedding_func OpenAIEmbeddings: Function used to embed the text

        connection_args Optional dict: args to connect Vector Database

        embedding_collection_name Optional str: exposed for test

        metadata_collection_name Optional str: exposed for test
    """
    self._embedding_func = embedding_func
    self._col_embeddings: Optional[Milvus] = None
    self._col_metadata: Optional[Collection] = None
    self._connection_args = validate_connection_args(connection_args)
    self._embedding_name = (
        embedding_collection_name or DOCUMENT_EMBEDDINGS_COLLECTION_NAME
    )
    self._metadata_name = (
        metadata_collection_name or DOCUMENT_METADATA_COLLECTION_NAME
    )

connect()

Connect to a host and read two document collections (the default names are DocumentEmbeddings and DocumentMetadata) in the currently active database (default database name is default); if those document collections don't exist, create the two collections.

Source code in biochatter/vectorstore_agent.py

def connect(self) -> None:
    """
    Connect to a host and read two document collections (the default names
    are `DocumentEmbeddings` and `DocumentMetadata`) in the currently active
    database (default database name is `default`); if those document
    collections don't exist, create the two collections.
    """
    self._connect(**self._connection_args)
    self._init_host()

get_all_documents(doc_ids=None)

Get all non-deleted documents from the currently active database.

Parameters:

Name	Type	Description	Default
doc_ids(List[str]	optional	the list of document ids, defines documents scope within	required

Returns:

Type	Description
list[dict]	List[Dict]: the metadata of all non-deleted documents in the form [{{id}, {author}, {source}, ...}]

Source code in biochatter/vectorstore_agent.py

def get_all_documents(
    self, doc_ids: Optional[list[str]] = None
) -> list[dict]:
    """
    Get all non-deleted documents from the currently active database.

    Args:
        doc_ids(List[str] optional): the list of document ids, defines documents scope within
        which the operation of obaining all documents occurs

    Returns:
        List[Dict]: the metadata of all non-deleted documents in the form
            [{{id}, {author}, {source}, ...}]
    """
    try:
        expr = VectorDatabaseAgentMilvus._build_meta_col_query_expr_for_all_documents(
            doc_ids
        )
        result_metadata = self._col_metadata.query(
            expr=expr, output_fields=METADATA_FIELDS
        )
        return result_metadata
    except MilvusException as e:
        logger.error(e)
        raise e

remove_document(doc_id, doc_ids=None)

Remove the document include meta data and its embeddings.

Parameters:

Name	Type	Description	Default
doc_id	str	the document to be deleted	required
doc_ids(List[str]	optional	the list of document ids, defines documents scope	required

Returns:

Name	Type	Description
bool	bool	True if the document is deleted, False otherwise

Source code in biochatter/vectorstore_agent.py

def remove_document(
    self, doc_id: str, doc_ids: Optional[list[str]] = None
) -> bool:
    """
    Remove the document include meta data and its embeddings.

    Args:
        doc_id (str): the document to be deleted

        doc_ids(List[str] optional): the list of document ids, defines documents scope
        within which remove operation occurs.

    Returns:
        bool: True if the document is deleted, False otherwise
    """
    if not self._col_metadata:
        return False
    if doc_ids is not None and (
        len(doc_ids) == 0 or (len(doc_ids) > 0 and not doc_id in doc_ids)
    ):
        return False
    try:
        expr = f"id in [{doc_id}]"
        res = self._col_metadata.query(
            expr=expr, output_fields=METADATA_FIELDS
        )
        if len(res) == 0:
            return False
        del_res = self._col_metadata.delete(expr)
        self._col_metadata.flush()

        res = self._col_embeddings.col.query(f'meta_id in ["{doc_id}"]')
        if len(res) == 0:
            return True
        ids = [item["pk"] for item in res]
        embedding_expr = f"pk in {ids}"
        del_res = self._col_embeddings.col.delete(expr=embedding_expr)
        self._col_embeddings.col.flush()
        return True
    except MilvusException as e:
        logger.error(e)
        raise e

similarity_search(query, k=3, doc_ids=None)

Perform similarity search insider the currently active database according to the input query.

This method will: 1. get all non-deleted meta_id and build to search expression for the currently active embedding collection 2. do similarity search in the embedding collection 3. combine metadata and embeddings

Parameters:

Name	Type	Description	Default
query	str	query string	required
k	int	the number of results to return	3
doc_ids(List[str]	optional	the list of document ids, do similarity search across the	required

Returns:

Type	Description
list[Document]	List[Document]: search results

Source code in biochatter/vectorstore_agent.py

def similarity_search(
    self, query: str, k: int = 3, doc_ids: list[str] = None
) -> list[Document]:
    """
    Perform similarity search insider the currently active database
    according to the input query.

    This method will:
    1. get all non-deleted meta_id and build to search expression for
        the currently active embedding collection
    2. do similarity search in the embedding collection
    3. combine metadata and embeddings

    Args:
        query (str): query string

        k (int): the number of results to return

        doc_ids(List[str] optional): the list of document ids, do similarity search across the
        specified documents

    Returns:
        List[Document]: search results
    """
    result_metadata = []
    expr = VectorDatabaseAgentMilvus._build_meta_col_query_expr_for_all_documents(
        doc_ids
    )
    result_metadata = self._col_metadata.query(
        expr=expr, output_fields=METADATA_FIELDS
    )
    expr = self._build_embedding_search_expression(result_metadata)
    result_embedding = self._col_embeddings.similarity_search(
        query=query, k=k, expr=expr
    )
    return self._join_embedding_and_metadata_results(
        result_embedding, result_metadata
    )

store_embeddings(documents)

Store documents in the currently active database.

Parameters:

Name	Type	Description	Default
documents	List[Documents]	documents array, usually from DocumentReader.load_document, DocumentReader.document_from_pdf, DocumentReader.document_from_txt	required

Returns:

Name	Type	Description
str	str	document id

Source code in biochatter/vectorstore_agent.py

def store_embeddings(self, documents: list[Document]) -> str:
    """
    Store documents in the currently active database.

    Args:
        documents (List[Documents]): documents array, usually from
            DocumentReader.load_document, DocumentReader.document_from_pdf,
            DocumentReader.document_from_txt

    Returns:
        str: document id
    """
    if len(documents) == 0:
        return
    return self._insert_data(documents)

align_embeddings(docs, meta_id)

Ensure that the metadata id is present in each document.

Parameters:

Name	Type	Description	Default
docs	List[Document]	List of documents	required
meta_id	int	Metadata id to assign to the documents	required

Returns:

Type	Description
list[Document]	List[Document]: List of documents, with each document having a metadata id.

Source code in biochatter/vectorstore_agent.py

def align_embeddings(docs: list[Document], meta_id: int) -> list[Document]:
    """
    Ensure that the metadata id is present in each document.

    Args:
        docs (List[Document]): List of documents

        meta_id (int): Metadata id to assign to the documents

    Returns:
        List[Document]: List of documents, with each document having a metadata
            id.
    """
    ret = []
    for doc in docs:
        ret.append(
            Document(
                page_content=doc.page_content,
                metadata={"meta_id": meta_id},
            )
        )
    return ret

align_metadata(metadata, isDeleted=False)

Ensure that specific metadata fields are present; if not provided, fill with "unknown". Also, add a random vector to each metadata item to simulate an embedding.

Parameters:

Name	Type	Description	Default
metadata	List[Dict]	List of metadata items	required
isDeleted	Optional[bool]	Whether the document is deleted. Defaults to False.	False

Returns:

Type	Description
list[list]	List[List]: List of metadata items, with each item being a list of metadata fields.

Source code in biochatter/vectorstore_agent.py

def align_metadata(
    metadata: list[dict], isDeleted: Optional[bool] = False
) -> list[list]:
    """

    Ensure that specific metadata fields are present; if not provided, fill with
    "unknown". Also, add a random vector to each metadata item to simulate an
    embedding.

    Args:
        metadata (List[Dict]): List of metadata items

        isDeleted (Optional[bool], optional): Whether the document is deleted.
            Defaults to False.

    Returns:
        List[List]: List of metadata items, with each item being a list of
            metadata fields.
    """
    ret = []
    fields = METADATA_FIELDS.copy()
    fields.pop(0)
    for ix, k in enumerate(fields):
        ret.append([item[k] if k in item else "unknown" for item in metadata])

    ret.append(
        [
            [random.random() for _ in range(METADATA_VECTOR_DIM)]
            for _ in range(len(metadata))
        ]
    )
    ret.append([isDeleted for _ in metadata])
    return ret