Redis

RedisVectorStore

基类: BasePydanticVectorStore

RedisVectorStore。

RedisVectorStore 接受用户定义的 schema 对象以及 Redis 连接客户端或 URL 字符串。 schema 是可选的，但对于以下情况很有用： - 定义自定义索引名称、键前缀和键分隔符。 - 定义 *额外* 元数据字段以用作查询过滤器。 - 在字段上设置自定义规范以提高搜索质量，例如使用哪种向量索引算法。

其他注意事项： - 所有嵌入和文档都存储在 Redis 中。在查询时，索引使用 Redis 查询最相似的前 k 个节点。 - Redis & LlamaIndex 要求任何 schema（默认或自定义）都至少有 4 个 *必需* 字段：id、doc_id、text、vector。

参数

名称	类型	描述	默认值
schema	IndexSchema	Redis 索引 schema 对象。	无
redis_client	Redis	Redis 客户端连接。	无
redis_url	str	Redis 服务器 URL。默认为 "redis://localhost:6379"。	无
overwrite	bool	如果索引已存在，是否覆盖。默认为 False。	False

引发

类型	描述
ValueError	如果您的 Redis 服务器未启用搜索或 JSON。
ValueError	如果 Redis 连接未能建立。
ValueError	如果提供了无效的 schema。

示例

from redisvl.schema import IndexSchema from llama_index.vector_stores.redis import RedisVectorStore

使用默认 schema

rds = RedisVectorStore(redis_url="redis://localhost:6379")

从 dict 使用自定义 schema

schema = IndexSchema.from_dict({ "index": {"name": "my-index", "prefix": "docs"}, "fields": [ {"name": "id", "type": "tag"}, {"name": "doc_id", "type": "tag"}, {"name": "text", "type": "text"}, {"name": "vector", "type": "vector", "attrs": {"dims": 1536, "algorithm": "flat"}} ] }) vector_store = RedisVectorStore( schema=schema, redis_url="redis://localhost:6379" )

源代码位于 llama-index-integrations/vector_stores/llama-index-vector-stores-redis/llama_index/vector_stores/redis/base.py

class RedisVectorStore(BasePydanticVectorStore):
    """
    RedisVectorStore.

    The RedisVectorStore takes a user-defined schema object and a Redis connection
    client or URL string. The schema is optional, but useful for:
    - Defining a custom index name, key prefix, and key separator.
    - Defining *additional* metadata fields to use as query filters.
    - Setting custom specifications on fields to improve search quality, e.g
    which vector index algorithm to use.

    Other Notes:
    - All embeddings and docs are stored in Redis. During query time, the index
    uses Redis to query for the top k most similar nodes.
    - Redis & LlamaIndex expect at least 4 *required* fields for any schema, default or custom,
    `id`, `doc_id`, `text`, `vector`.

    Args:
        schema (IndexSchema, optional): Redis index schema object.
        redis_client (Redis, optional): Redis client connection.
        redis_url (str, optional): Redis server URL.
            Defaults to "redis://localhost:6379".
        overwrite (bool, optional): Whether to overwrite the index if it already exists.
            Defaults to False.

    Raises:
        ValueError: If your Redis server does not have search or JSON enabled.
        ValueError: If a Redis connection failed to be established.
        ValueError: If an invalid schema is provided.

    Example:
        from redisvl.schema import IndexSchema
        from llama_index.vector_stores.redis import RedisVectorStore

        # Use default schema
        rds = RedisVectorStore(redis_url="redis://localhost:6379")

        # Use custom schema from dict
        schema = IndexSchema.from_dict({
            "index": {"name": "my-index", "prefix": "docs"},
            "fields": [
                {"name": "id", "type": "tag"},
                {"name": "doc_id", "type": "tag},
                {"name": "text", "type": "text"},
                {"name": "vector", "type": "vector", "attrs": {"dims": 1536, "algorithm": "flat"}}
            ]
        })
        vector_store = RedisVectorStore(
            schema=schema,
            redis_url="redis://localhost:6379"
        )

    """

    stores_text: bool = True
    stores_node: bool = True
    flat_metadata: bool = False

    _index: SearchIndex = PrivateAttr()
    _overwrite: bool = PrivateAttr()
    _return_fields: List[str] = PrivateAttr()

    def __init__(
        self,
        schema: Optional[IndexSchema] = None,
        redis_client: Optional[Redis] = None,
        redis_url: Optional[str] = None,
        overwrite: bool = False,
        return_fields: Optional[List[str]] = None,
        **kwargs: Any,
    ) -> None:
        super().__init__()
        # check for indicators of old schema
        self._flag_old_kwargs(**kwargs)

        # Setup schema
        if not schema:
            logger.info("Using default RedisVectorStore schema.")
            schema = RedisVectorStoreSchema()

        self._validate_schema(schema)
        self._return_fields = return_fields or [
            NODE_ID_FIELD_NAME,
            DOC_ID_FIELD_NAME,
            TEXT_FIELD_NAME,
            NODE_CONTENT_FIELD_NAME,
        ]
        self._index = SearchIndex(
            schema=schema, redis_client=redis_client, redis_url=redis_url
        )
        self._overwrite = overwrite

        # Create index
        self.create_index()

    def _flag_old_kwargs(self, **kwargs):
        old_kwargs = [
            "index_name",
            "index_prefix",
            "prefix_ending",
            "index_args",
            "metadata_fields",
        ]
        for kwarg in old_kwargs:
            if kwarg in kwargs:
                raise ValueError(
                    f"Deprecated kwarg, {kwarg}, found upon initialization. "
                    "RedisVectorStore now requires an IndexSchema object. "
                    "See the documentation for a complete example: https://docs.llamaindex.org.cn/en/stable/examples/vector_stores/RedisIndexDemo/"
                )

    def _validate_schema(self, schema: IndexSchema) -> str:
        base_schema = RedisVectorStoreSchema()
        for name, field in base_schema.fields.items():
            if (name not in schema.fields) or (
                not schema.fields[name].type == field.type
            ):
                raise ValueError(
                    f"Required field {name} must be present in the index "
                    f"and of type {schema.fields[name].type}"
                )

    @property
    def client(self) -> "Redis":
        """Return the redis client instance."""
        return self._index.client

    @property
    def index_name(self) -> str:
        """Return the name of the index based on the schema."""
        return self._index.name

    @property
    def schema(self) -> IndexSchema:
        """Return the index schema."""
        return self._index.schema

    def set_return_fields(self, return_fields: List[str]) -> None:
        """Update the return fields for the query response."""
        self._return_fields = return_fields

    def index_exists(self) -> bool:
        """
        Check whether the index exists in Redis.

        Returns:
            bool: True or False.

        """
        return self._index.exists()

    def create_index(self, overwrite: Optional[bool] = None) -> None:
        """Create an index in Redis."""
        if overwrite is None:
            overwrite = self._overwrite
        # Create index honoring overwrite policy
        if overwrite:
            self._index.create(overwrite=True, drop=True)
        else:
            self._index.create()

    def add(self, nodes: List[BaseNode], **add_kwargs: Any) -> List[str]:
        """
        Add nodes to the index.

        Args:
            nodes (List[BaseNode]): List of nodes with embeddings

        Returns:
            List[str]: List of ids of the documents added to the index.

        Raises:
            ValueError: If the index already exists and overwrite is False.

        """
        # Check to see if empty document list was passed
        if len(nodes) == 0:
            return []

        # Now check for the scenario where user is trying to index embeddings that don't align with schema
        embedding_len = len(nodes[0].get_embedding())
        expected_dims = self._index.schema.fields[VECTOR_FIELD_NAME].attrs.dims
        if expected_dims != embedding_len:
            raise ValueError(
                f"Attempting to index embeddings of dim {embedding_len} "
                f"which doesn't match the index schema expectation of {expected_dims}. "
                "Please review the Redis integration example to learn how to customize schema. "
                ""
            )

        data: List[Dict[str, Any]] = []
        for node in nodes:
            embedding = node.get_embedding()
            record = {
                NODE_ID_FIELD_NAME: node.node_id,
                DOC_ID_FIELD_NAME: node.ref_doc_id,
                TEXT_FIELD_NAME: node.get_content(metadata_mode=MetadataMode.NONE),
                VECTOR_FIELD_NAME: array_to_buffer(embedding, dtype="FLOAT32"),
            }
            # parse and append metadata
            additional_metadata = node_to_metadata_dict(
                node, remove_text=True, flat_metadata=self.flat_metadata
            )
            data.append({**record, **additional_metadata})

        # Load nodes to Redis
        keys = self._index.load(data, id_field=NODE_ID_FIELD_NAME, **add_kwargs)
        logger.info(f"Added {len(keys)} documents to index {self._index.name}")
        return [
            key.strip(self._index.prefix + self._index.key_separator) for key in keys
        ]

    def delete(self, ref_doc_id: str, **delete_kwargs: Any) -> None:
        """
        Delete nodes using with ref_doc_id.

        Args:
            ref_doc_id (str): The doc_id of the document to delete.

        """
        # build a filter to target specific docs by doc ID
        doc_filter = Tag(DOC_ID_FIELD_NAME) == ref_doc_id
        total = self._index.query(CountQuery(doc_filter))
        delete_query = FilterQuery(
            return_fields=[NODE_ID_FIELD_NAME],
            filter_expression=doc_filter,
            num_results=total,
        )
        # fetch docs to delete and flush them
        docs_to_delete = self._index.search(delete_query.query, delete_query.params)
        with self._index.client.pipeline(transaction=False) as pipe:
            for doc in docs_to_delete.docs:
                pipe.delete(doc.id)
            res = pipe.execute()

        logger.info(
            f"Deleted {len(docs_to_delete.docs)} documents from index {self._index.name}"
        )

    def delete_index(self) -> None:
        """Delete the index and all documents."""
        logger.info(f"Deleting index {self._index.name}")
        self._index.delete(drop=True)

    @staticmethod
    def _to_redis_filter(field: BaseField, filter: MetadataFilter) -> FilterExpression:
        """
        Translate a standard metadata filter to a Redis specific filter expression.

        Args:
            field (BaseField): The field to be filtered on, must have a type attribute.
            filter (MetadataFilter): The filter to apply, must have operator and value attributes.

        Returns:
            FilterExpression: A Redis-specific filter expression constructed from the input.

        Raises:
            ValueError: If the field type is unsupported or if the operator is not supported for the field type.

        """
        # Check for unsupported field type
        if field.type not in REDIS_LLAMA_FIELD_SPEC:
            raise ValueError(f"Unsupported field type {field.type} for {field.name}")

        field_info = REDIS_LLAMA_FIELD_SPEC[field.type]

        # Check for unsupported operator
        if filter.operator not in field_info["operators"]:
            raise ValueError(
                f"Filter operator {filter.operator} not supported for {field.name} of type {field.type}"
            )

        # Create field instance and apply the operator function
        field_instance = field_info["class"](field.name)
        return field_info["operators"][filter.operator](field_instance, filter.value)

    def _create_redis_filter_expression(
        self, metadata_filters: MetadataFilters
    ) -> FilterExpression:
        """
        Generate a Redis Filter Expression as a combination of metadata filters.

        Args:
            metadata_filters (MetadataFilters): List of metadata filters to use.

        Returns:
            FilterExpression: A Redis filter expression.

        """
        filter_expression = FilterExpression("*")
        if metadata_filters:
            if metadata_filters.filters:
                for filter in metadata_filters.filters:
                    # Handle nested MetadataFilters recursively
                    if isinstance(filter, MetadataFilters):
                        redis_filter = self._create_redis_filter_expression(filter)
                    else:
                        # Index must be created with the metadata field in the index schema
                        field = self._index.schema.fields.get(filter.key)
                        if not field:
                            logger.warning(
                                f"{filter.key} field was not included as part of the index schema, and thus cannot be used as a filter condition."
                            )
                            continue
                        # Extract redis filter
                        redis_filter = self._to_redis_filter(field, filter)

                    # Combine with conditional
                    if metadata_filters.condition == "and":
                        filter_expression = filter_expression & redis_filter
                    else:
                        filter_expression = filter_expression | redis_filter
        return filter_expression

    def _to_redis_query(self, query: VectorStoreQuery) -> VectorQuery:
        """Creates a RedisQuery from a VectorStoreQuery."""
        filter_expression = self._create_redis_filter_expression(query.filters)
        return_fields = self._return_fields.copy()
        return VectorQuery(
            vector=query.query_embedding,
            vector_field_name=VECTOR_FIELD_NAME,
            num_results=query.similarity_top_k,
            filter_expression=filter_expression,
            return_fields=return_fields,
        )

    def _extract_node_and_score(self, doc, redis_query: VectorQuery):
        """Extracts a node and its score from a document."""
        try:
            node = metadata_dict_to_node(
                {NODE_CONTENT_FIELD_NAME: doc[NODE_CONTENT_FIELD_NAME]}
            )
            node.text = doc[TEXT_FIELD_NAME]
        except Exception:
            # Handle legacy metadata format
            node = TextNode(
                text=doc[TEXT_FIELD_NAME],
                id_=doc[NODE_ID_FIELD_NAME],
                embedding=None,
                relationships={
                    NodeRelationship.SOURCE: RelatedNodeInfo(
                        node_id=doc[DOC_ID_FIELD_NAME]
                    )
                },
            )
        score = 1 - float(doc[redis_query.DISTANCE_ID])
        return node, score

    def _process_query_results(
        self, results, redis_query: VectorQuery
    ) -> VectorStoreQueryResult:
        """Processes query results and returns a VectorStoreQueryResult."""
        ids, nodes, scores = [], [], []
        for doc in results:
            node, score = self._extract_node_and_score(doc, redis_query)
            ids.append(doc[NODE_ID_FIELD_NAME])
            nodes.append(node)
            scores.append(score)
        logger.info(f"Found {len(nodes)} results for query with id {ids}")
        return VectorStoreQueryResult(nodes=nodes, ids=ids, similarities=scores)

    def query(self, query: VectorStoreQuery, **kwargs: Any) -> VectorStoreQueryResult:
        """
        Query the index.

        Args:
            query (VectorStoreQuery): query object

        Returns:
            VectorStoreQueryResult: query result

        Raises:
            ValueError: If query.query_embedding is None.
            redis.exceptions.RedisError: If there is an error querying the index.
            redis.exceptions.TimeoutError: If there is a timeout querying the index.

        """
        if not query.query_embedding:
            raise ValueError("Query embedding is required for querying.")

        redis_query = self._to_redis_query(query)
        logger.info(f"Querying index {self._index.name} with query {redis_query!s}")

        try:
            results = self._index.query(redis_query)
        except RedisTimeoutError as e:
            logger.error(f"Query timed out on {self._index.name}: {e}")
            raise
        except RedisError as e:
            logger.error(f"Error querying {self._index.name}: {e}")
            raise

        return self._process_query_results(results, redis_query)

    def persist(
        self,
        persist_path: Optional[str] = None,
        fs: Optional[fsspec.AbstractFileSystem] = None,
        in_background: bool = True,
    ) -> None:
        """
        Persist the vector store to disk.

        For Redis, more notes here: https://redis.ac.cn/docs/management/persistence/

        Args:
            persist_path (str): Path to persist the vector store to. (doesn't apply)
            in_background (bool, optional): Persist in background. Defaults to True.
            fs (fsspec.AbstractFileSystem, optional): Filesystem to persist to.
                (doesn't apply)

        Raises:
            redis.exceptions.RedisError: If there is an error
                                         persisting the index to disk.

        """
        try:
            if in_background:
                logger.info("Saving index to disk in background")
                self._index.client.bgsave()
            else:
                logger.info("Saving index to disk")
                self._index.client.save()

        except RedisError as e:
            logger.error(f"Error saving index to disk: {e}")
            raise

client 属性

client: Redis

返回 redis 客户端实例。

index_name 属性

index_name: str

根据 schema 返回索引名称。

schema 属性

schema: IndexSchema

返回索引 schema。

set_return_fields

set_return_fields(return_fields: List[str]) -> None

更新查询响应的返回字段。

源代码位于 llama-index-integrations/vector_stores/llama-index-vector-stores-redis/llama_index/vector_stores/redis/base.py

def set_return_fields(self, return_fields: List[str]) -> None:
    """Update the return fields for the query response."""
    self._return_fields = return_fields

index_exists

index_exists() -> bool

检查索引是否存在于 Redis 中。

返回

名称	类型	描述
bool	bool	True 或 False。

源代码位于 llama-index-integrations/vector_stores/llama-index-vector-stores-redis/llama_index/vector_stores/redis/base.py

def index_exists(self) -> bool:
    """
    Check whether the index exists in Redis.

    Returns:
        bool: True or False.

    """
    return self._index.exists()

create_index

create_index(overwrite: Optional[bool] = None) -> None

在 Redis 中创建索引。

源代码位于 llama-index-integrations/vector_stores/llama-index-vector-stores-redis/llama_index/vector_stores/redis/base.py

def create_index(self, overwrite: Optional[bool] = None) -> None:
    """Create an index in Redis."""
    if overwrite is None:
        overwrite = self._overwrite
    # Create index honoring overwrite policy
    if overwrite:
        self._index.create(overwrite=True, drop=True)
    else:
        self._index.create()

add

add(nodes: List[BaseNode], **add_kwargs: Any) -> List[str]

向索引添加节点。

参数

名称	类型	描述	默认值
nodes	List[BaseNode]	包含嵌入的节点列表	必需

返回

类型	描述
List[str]	List[str]：添加到索引的文档 ID 列表。

引发

类型	描述
ValueError	如果索引已存在且 overwrite 为 False。

源代码位于 llama-index-integrations/vector_stores/llama-index-vector-stores-redis/llama_index/vector_stores/redis/base.py

def add(self, nodes: List[BaseNode], **add_kwargs: Any) -> List[str]:
    """
    Add nodes to the index.

    Args:
        nodes (List[BaseNode]): List of nodes with embeddings

    Returns:
        List[str]: List of ids of the documents added to the index.

    Raises:
        ValueError: If the index already exists and overwrite is False.

    """
    # Check to see if empty document list was passed
    if len(nodes) == 0:
        return []

    # Now check for the scenario where user is trying to index embeddings that don't align with schema
    embedding_len = len(nodes[0].get_embedding())
    expected_dims = self._index.schema.fields[VECTOR_FIELD_NAME].attrs.dims
    if expected_dims != embedding_len:
        raise ValueError(
            f"Attempting to index embeddings of dim {embedding_len} "
            f"which doesn't match the index schema expectation of {expected_dims}. "
            "Please review the Redis integration example to learn how to customize schema. "
            ""
        )

    data: List[Dict[str, Any]] = []
    for node in nodes:
        embedding = node.get_embedding()
        record = {
            NODE_ID_FIELD_NAME: node.node_id,
            DOC_ID_FIELD_NAME: node.ref_doc_id,
            TEXT_FIELD_NAME: node.get_content(metadata_mode=MetadataMode.NONE),
            VECTOR_FIELD_NAME: array_to_buffer(embedding, dtype="FLOAT32"),
        }
        # parse and append metadata
        additional_metadata = node_to_metadata_dict(
            node, remove_text=True, flat_metadata=self.flat_metadata
        )
        data.append({**record, **additional_metadata})

    # Load nodes to Redis
    keys = self._index.load(data, id_field=NODE_ID_FIELD_NAME, **add_kwargs)
    logger.info(f"Added {len(keys)} documents to index {self._index.name}")
    return [
        key.strip(self._index.prefix + self._index.key_separator) for key in keys
    ]

delete

delete(ref_doc_id: str, **delete_kwargs: Any) -> None

使用 ref_doc_id 删除节点。

参数

名称	类型	描述	默认值
ref_doc_id	str	要删除的文档的 doc_id。	必需

源代码位于 llama-index-integrations/vector_stores/llama-index-vector-stores-redis/llama_index/vector_stores/redis/base.py

def delete(self, ref_doc_id: str, **delete_kwargs: Any) -> None:
    """
    Delete nodes using with ref_doc_id.

    Args:
        ref_doc_id (str): The doc_id of the document to delete.

    """
    # build a filter to target specific docs by doc ID
    doc_filter = Tag(DOC_ID_FIELD_NAME) == ref_doc_id
    total = self._index.query(CountQuery(doc_filter))
    delete_query = FilterQuery(
        return_fields=[NODE_ID_FIELD_NAME],
        filter_expression=doc_filter,
        num_results=total,
    )
    # fetch docs to delete and flush them
    docs_to_delete = self._index.search(delete_query.query, delete_query.params)
    with self._index.client.pipeline(transaction=False) as pipe:
        for doc in docs_to_delete.docs:
            pipe.delete(doc.id)
        res = pipe.execute()

    logger.info(
        f"Deleted {len(docs_to_delete.docs)} documents from index {self._index.name}"
    )

delete_index

delete_index() -> None

删除索引和所有文档。

源代码位于 llama-index-integrations/vector_stores/llama-index-vector-stores-redis/llama_index/vector_stores/redis/base.py

def delete_index(self) -> None:
    """Delete the index and all documents."""
    logger.info(f"Deleting index {self._index.name}")
    self._index.delete(drop=True)

query

query(query: VectorStoreQuery, **kwargs: Any) -> VectorStoreQueryResult

查询索引。

参数

名称	类型	描述	默认值
query	VectorStoreQuery	查询对象	必需

返回

名称	类型	描述
VectorStoreQueryResult	VectorStoreQueryResult	查询结果

引发

类型	描述
ValueError	如果 query.query_embedding 为 None。
RedisError	如果在查询索引时发生错误。
TimeoutError	如果在查询索引时发生超时。

源代码位于 llama-index-integrations/vector_stores/llama-index-vector-stores-redis/llama_index/vector_stores/redis/base.py

def query(self, query: VectorStoreQuery, **kwargs: Any) -> VectorStoreQueryResult:
    """
    Query the index.

    Args:
        query (VectorStoreQuery): query object

    Returns:
        VectorStoreQueryResult: query result

    Raises:
        ValueError: If query.query_embedding is None.
        redis.exceptions.RedisError: If there is an error querying the index.
        redis.exceptions.TimeoutError: If there is a timeout querying the index.

    """
    if not query.query_embedding:
        raise ValueError("Query embedding is required for querying.")

    redis_query = self._to_redis_query(query)
    logger.info(f"Querying index {self._index.name} with query {redis_query!s}")

    try:
        results = self._index.query(redis_query)
    except RedisTimeoutError as e:
        logger.error(f"Query timed out on {self._index.name}: {e}")
        raise
    except RedisError as e:
        logger.error(f"Error querying {self._index.name}: {e}")
        raise

    return self._process_query_results(results, redis_query)

persist

persist(persist_path: Optional[str] = None, fs: Optional[AbstractFileSystem] = None, in_background: bool = True) -> None

将向量存储持久化到磁盘。

对于 Redis，更多注意事项在此处：https://redis.ac.cn/docs/management/persistence/

参数

名称	类型	描述	默认值
persist_path	str	向量存储持久化的路径。（不适用）	无
in_background	bool	在后台持久化。默认为 True。	True
fs	AbstractFileSystem	用于持久化的文件系统。（不适用）	无

引发

类型	描述
RedisError	如果将索引持久化到磁盘时发生错误。

源代码位于 llama-index-integrations/vector_stores/llama-index-vector-stores-redis/llama_index/vector_stores/redis/base.py

def persist(
    self,
    persist_path: Optional[str] = None,
    fs: Optional[fsspec.AbstractFileSystem] = None,
    in_background: bool = True,
) -> None:
    """
    Persist the vector store to disk.

    For Redis, more notes here: https://redis.ac.cn/docs/management/persistence/

    Args:
        persist_path (str): Path to persist the vector store to. (doesn't apply)
        in_background (bool, optional): Persist in background. Defaults to True.
        fs (fsspec.AbstractFileSystem, optional): Filesystem to persist to.
            (doesn't apply)

    Raises:
        redis.exceptions.RedisError: If there is an error
                                     persisting the index to disk.

    """
    try:
        if in_background:
            logger.info("Saving index to disk in background")
            self._index.client.bgsave()
        else:
            logger.info("Saving index to disk")
            self._index.client.save()

    except RedisError as e:
        logger.error(f"Error saving index to disk: {e}")
        raise