文档数组

class DocArrayHnswVectorStore(DocArrayVectorStore):
    """
    Class representing a DocArray HNSW vector store.

    This class is a lightweight Document Index implementation provided by Docarray.
    It stores vectors on disk in hnswlib, and stores all other data in SQLite.

    Examples:
        `pip install llama-index-vector-stores-docarray`

        ```python
        from llama_index.vector_stores.docarray import DocArrayHnswVectorStore

        # Initialize the DocArrayHnswVectorStore
        vector_store = DocArrayHnswVectorStore(work_dir="hnsw_index", dim=1536)
        ```

    """

    def __init__(
        self,
        work_dir: str,
        dim: int = 1536,
        dist_metric: Literal["cosine", "ip", "l2"] = "cosine",
        max_elements: int = 1024,
        ef_construction: int = 200,
        ef: int = 10,
        M: int = 16,
        allow_replace_deleted: bool = True,
        num_threads: int = 1,
    ):
        """
        Initializes the DocArrayHnswVectorStore.

        Args:
            work_dir (str): The working directory.
            dim (int, optional): Dimensionality of the vectors. Default is 1536.
            dist_metric (Literal["cosine", "ip", "l2"], optional): The distance
                metric to use. Default is "cosine".
            max_elements (int, optional): defines the maximum number of elements
                that can be stored in the structure(can be increased/shrunk).
            ef_construction (int, optional): defines a construction time/accuracy
                trade-off. Default is 200.
            ef (int, optional): The size of the dynamic candidate list. Default is 10.
            M (int, optional): defines the maximum number of outgoing connections
                in the graph. Default is 16.
            allow_replace_deleted (bool, optional): Whether to allow replacing
                deleted elements. Default is True.
            num_threads (int, optional): Number of threads for index construction.
                Default is 1.

        """
        import_err_msg = """
                `docarray` package not found. Install the package via pip:
                `pip install docarray[hnswlib]`
        """
        try:
            import docarray  # noqa
        except ImportError:
            raise ImportError(import_err_msg)

        self._work_dir = work_dir
        ref_docs_path = os.path.join(self._work_dir, "ref_docs.json")
        if os.path.exists(ref_docs_path):
            with open(ref_docs_path) as f:
                self._ref_docs = json.load(f)
        else:
            self._ref_docs = {}

        self._index, self._schema = self._init_index(
            dim=dim,
            dist_metric=dist_metric,
            max_elements=max_elements,
            ef_construction=ef_construction,
            ef=ef,
            M=M,
            allow_replace_deleted=allow_replace_deleted,
            num_threads=num_threads,
        )

    def _init_index(self, **kwargs: Any):  # type: ignore[no-untyped-def]
        """
        Initializes the HNSW document index.

        Args:
            **kwargs: Variable length argument list for the HNSW index.

        Returns:
            tuple: The HNSW document index and its schema.

        """
        from docarray.index import HnswDocumentIndex

        schema = self._get_schema(**kwargs)
        index = HnswDocumentIndex[schema]  # type: ignore[valid-type]
        return index(work_dir=self._work_dir), schema

    def _find_docs_to_be_removed(self, doc_id: str) -> List[str]:
        """
        Finds the documents to be removed from the vector store.

        Args:
            doc_id (str): Reference document ID that should be removed.

        Returns:
            List[str]: List of document IDs to be removed.

        """
        docs = self._ref_docs.get(doc_id, [])
        del self._ref_docs[doc_id]
        self._save_ref_docs()
        return docs

    def _save_ref_docs(self) -> None:
        """Saves reference documents."""
        with open(os.path.join(self._work_dir, "ref_docs.json"), "w") as f:
            json.dump(self._ref_docs, f)

    def _update_ref_docs(self, docs):  # type: ignore[no-untyped-def]
        """
        Updates reference documents.

        Args:
            docs (List): List of documents to update.

        """
        for doc in docs:
            if doc.metadata["doc_id"] not in self._ref_docs:
                self._ref_docs[doc.metadata["doc_id"]] = []
            self._ref_docs[doc.metadata["doc_id"]].append(doc.id)
        self._save_ref_docs()

class DocArrayInMemoryVectorStore(DocArrayVectorStore):
    """
    Class representing a DocArray In-Memory vector store.

    This class is a document index provided by Docarray that stores documents in memory.

    Examples:
        `pip install llama-index-vector-stores-docarray`

        ```python
        from llama_index.vector_stores.docarray import DocArrayInMemoryVectorStore

        # Create an instance of DocArrayInMemoryVectorStore
        vector_store = DocArrayInMemoryVectorStore()
        ```

    """

    def __init__(
        self,
        index_path: Optional[str] = None,
        metric: Literal[
            "cosine_sim", "euclidian_dist", "sgeuclidean_dist"
        ] = "cosine_sim",
    ):
        """
        Initializes the DocArrayInMemoryVectorStore.

        Args:
            index_path (Optional[str]): The path to the index file.
            metric (Literal["cosine_sim", "euclidian_dist", "sgeuclidean_dist"]):
                The distance metric to use. Default is "cosine_sim".

        """
        import_err_msg = """
                `docarray` package not found. Install the package via pip:
                `pip install docarray`
        """
        try:
            import docarray  # noqa
        except ImportError:
            raise ImportError(import_err_msg)

        self._ref_docs = None  # type: ignore[assignment]
        self._index_file_path = index_path
        self._index, self._schema = self._init_index(metric=metric)

    def _init_index(self, **kwargs: Any):  # type: ignore[no-untyped-def]
        """
        Initializes the in-memory exact nearest neighbour index.

        Args:
            **kwargs: Variable length argument list.

        Returns:
            tuple: The in-memory exact nearest neighbour index and its schema.

        """
        from docarray.index import InMemoryExactNNIndex

        schema = self._get_schema(**kwargs)
        index = InMemoryExactNNIndex[schema]  # type: ignore[valid-type]
        params = {"index_file_path": self._index_file_path}
        return index(**params), schema  # type: ignore[arg-type]

    def _find_docs_to_be_removed(self, doc_id: str) -> List[str]:
        """
        Finds the documents to be removed from the vector store.

        Args:
            doc_id (str): Reference document ID that should be removed.

        Returns:
            List[str]: List of document IDs to be removed.

        """
        query = {"metadata__doc_id": {"$eq": doc_id}}
        docs = self._index.filter(query)
        return [doc.id for doc in docs]

    def persist(
        self, persist_path: str, fs: Optional[fsspec.AbstractFileSystem] = None
    ) -> None:
        """
        Persists the in-memory vector store to a file.

        Args:
            persist_path (str): The path to persist the index.
            fs (fsspec.AbstractFileSystem, optional): Filesystem to persist to.
                (doesn't apply)

        """
        index_path = persist_path or self._index_file_path
        self._index.persist(index_path)