from __future__ import annotations import json import logging import os import warnings from dataclasses import dataclass, field from hashlib import md5 from typing import Any, Iterable, Iterator, List, Optional, Tuple, Type import requests from langchain_core.callbacks.manager import ( CallbackManagerForRetrieverRun, ) from langchain_core.documents import Document from langchain_core.embeddings import Embeddings from langchain_core.runnables import Runnable, RunnableConfig from langchain_core.vectorstores import VectorStore, VectorStoreRetriever from pydantic import ConfigDict logger = logging.getLogger(__name__) MMR_RERANKER_ID = 272725718 RERANKER_MULTILINGUAL_V1_ID = 272725719 UDF_RERANKER_ID = 272725722 @dataclass class SummaryConfig: """Configuration for summary generation. is_enabled: True if summary is enabled, False otherwise max_results: maximum number of results to summarize response_lang: requested language for the summary prompt_name: name of the prompt to use for summarization (see https://docs.vectara.com/docs/learn/grounded-generation/select-a-summarizer) """ is_enabled: bool = False max_results: int = 7 response_lang: str = "eng" prompt_name: str = "vectara-summary-ext-24-05-med-omni" stream: bool = False @dataclass class MMRConfig: """Configuration for Maximal Marginal Relevance (MMR) search. This will soon be deprated in favor of RerankConfig. is_enabled: True if MMR is enabled, False otherwise mmr_k: number of results to fetch for MMR, defaults to 50 diversity_bias: number between 0 and 1 that determines the degree of diversity among the results with 0 corresponding to minimum diversity and 1 to maximum diversity. Defaults to 0.3. Note: diversity_bias is equivalent 1-lambda_mult where lambda_mult is the value often used in max_marginal_relevance_search() We chose to use that since we believe it's more intuitive to the user. """ is_enabled: bool = False mmr_k: int = 50 diversity_bias: float = 0.3 @dataclass class RerankConfig: """Configuration for Reranker. reranker: "mmr", "rerank_multilingual_v1", "udf" or "none" rerank_k: number of results to fetch before reranking, defaults to 50 mmr_diversity_bias: for MMR only - a number between 0 and 1 that determines the degree of diversity among the results with 0 corresponding to minimum diversity and 1 to maximum diversity. Defaults to 0.3. Note: mmr_diversity_bias is equivalent 1-lambda_mult where lambda_mult is the value often used in max_marginal_relevance_search() We chose to use that since we believe it's more intuitive to the user. user_function: for UDF only - the user function to use for reranking. """ reranker: str = "none" rerank_k: int = 50 mmr_diversity_bias: float = 0.3 user_function: str = "" @dataclass class VectaraQueryConfig: """Configuration for Vectara query. k: Number of Documents to return. Defaults to 10. lambda_val: lexical match parameter for hybrid search. filter Dictionary of argument(s) to filter on metadata. For example a filter can be "doc.rating > 3.0 and part.lang = 'deu'"} see https://docs.vectara.com/docs/search-apis/sql/filter-overview for more details. score_threshold: minimal score threshold for the result. If defined, results with score less than this value will be filtered out. n_sentence_before: number of sentences before the matching segment to add, defaults to 2 n_sentence_after: number of sentences before the matching segment to add, defaults to 2 rerank_config: RerankConfig configuration dataclass summary_config: SummaryConfig configuration dataclass """ k: int = 10 lambda_val: float = 0.0 filter: str = "" score_threshold: Optional[float] = None n_sentence_before: int = 2 n_sentence_after: int = 2 rerank_config: RerankConfig = field(default_factory=RerankConfig) summary_config: SummaryConfig = field(default_factory=SummaryConfig) def __init__( self, k: int = 10, lambda_val: float = 0.0, filter: str = "", score_threshold: Optional[float] = None, n_sentence_before: int = 2, n_sentence_after: int = 2, n_sentence_context: Optional[int] = None, mmr_config: Optional[MMRConfig] = None, summary_config: Optional[SummaryConfig] = None, rerank_config: Optional[RerankConfig] = None, ): self.k = k self.lambda_val = lambda_val self.filter = filter self.score_threshold = score_threshold if summary_config: self.summary_config = summary_config else: self.summary_config = SummaryConfig() # handle n_sentence_context for backward compatibility if n_sentence_context: self.n_sentence_before = n_sentence_context self.n_sentence_after = n_sentence_context warnings.warn( "n_sentence_context is deprecated. " "Please use n_sentence_before and n_sentence_after instead", DeprecationWarning, ) else: self.n_sentence_before = n_sentence_before self.n_sentence_after = n_sentence_after # handle mmr_config for backward compatibility if rerank_config: self.rerank_config = rerank_config elif mmr_config: self.rerank_config = RerankConfig( reranker="mmr", rerank_k=mmr_config.mmr_k, mmr_diversity_bias=mmr_config.diversity_bias, ) warnings.warn( "MMRConfig is deprecated. Please use RerankConfig instead.", DeprecationWarning, ) else: self.rerank_config = RerankConfig() class Vectara(VectorStore): """`Vectara API` vector store. See (https://vectara.com). Example: .. code-block:: python from langchain_community.vectorstores import Vectara vectorstore = Vectara( vectara_customer_id=vectara_customer_id, vectara_corpus_id=vectara_corpus_id, vectara_api_key=vectara_api_key ) """ def __init__( self, vectara_customer_id: Optional[str] = None, vectara_corpus_id: Optional[str] = None, vectara_api_key: Optional[str] = None, vectara_api_timeout: int = 120, source: str = "langchain", ): """Initialize with Vectara API.""" self._vectara_customer_id = vectara_customer_id or os.environ.get( "VECTARA_CUSTOMER_ID" ) self._vectara_corpus_id = vectara_corpus_id or os.environ.get( "VECTARA_CORPUS_ID" ) self._vectara_api_key = vectara_api_key or os.environ.get("VECTARA_API_KEY") if ( self._vectara_customer_id is None or self._vectara_corpus_id is None or self._vectara_api_key is None ): logger.warning( "Can't find Vectara credentials, customer_id or corpus_id in " "environment." ) else: logger.debug(f"Using corpus id {self._vectara_corpus_id}") self._source = source self._session = requests.Session() # to reuse connections adapter = requests.adapters.HTTPAdapter(max_retries=3) self._session.mount("http://", adapter) self.vectara_api_timeout = vectara_api_timeout @property def embeddings(self) -> Optional[Embeddings]: return None def _get_post_headers(self) -> dict: """Returns headers that should be attached to each post request.""" return { "x-api-key": self._vectara_api_key, "customer-id": self._vectara_customer_id, "Content-Type": "application/json", "X-Source": self._source, } def _delete_doc(self, doc_id: str) -> bool: """ Delete a document from the Vectara corpus. Args: doc_id (str): ID of the document to delete. Returns: bool: True if deletion was successful, False otherwise. """ body = { "customer_id": self._vectara_customer_id, "corpus_id": self._vectara_corpus_id, "document_id": doc_id, } response = self._session.post( "https://api.vectara.io/v1/delete-doc", data=json.dumps(body), verify=True, headers=self._get_post_headers(), timeout=self.vectara_api_timeout, ) if response.status_code != 200: logger.error( f"Delete request failed for doc_id = {doc_id} with status code " f"{response.status_code}, reason {response.reason}, text " f"{response.text}" ) return False return True def _index_doc(self, doc: dict, use_core_api: bool = False) -> str: request: dict[str, Any] = {} request["customer_id"] = self._vectara_customer_id request["corpus_id"] = self._vectara_corpus_id request["document"] = doc api_endpoint = ( "https://api.vectara.io/v1/core/index" if use_core_api else "https://api.vectara.io/v1/index" ) response = self._session.post( headers=self._get_post_headers(), url=api_endpoint, data=json.dumps(request), timeout=self.vectara_api_timeout, verify=True, ) status_code = response.status_code result = response.json() status_str = result["status"]["code"] if "status" in result else None if status_code == 409 or status_str and (status_str == "ALREADY_EXISTS"): return "E_ALREADY_EXISTS" elif status_str and (status_str == "FORBIDDEN"): return "E_NO_PERMISSIONS" else: return "E_SUCCEEDED" def delete(self, ids: Optional[List[str]] = None, **kwargs: Any) -> Optional[bool]: """Delete by vector ID or other criteria. Args: ids: List of ids to delete. Returns: Optional[bool]: True if deletion is successful, False otherwise, None if not implemented. """ if ids: success = [self._delete_doc(id) for id in ids] return all(success) else: return True def add_files( self, files_list: Iterable[str], metadatas: Optional[List[dict]] = None, **kwargs: Any, ) -> List[str]: """ Vectara provides a way to add documents directly via our API where pre-processing and chunking occurs internally in an optimal way This method provides a way to use that API in LangChain Args: files_list: Iterable of strings, each representing a local file path. Files could be text, HTML, PDF, markdown, doc/docx, ppt/pptx, etc. see API docs for full list metadatas: Optional list of metadatas associated with each file Returns: List of ids associated with each of the files indexed """ doc_ids = [] for inx, file in enumerate(files_list): if not os.path.exists(file): logger.error(f"File {file} does not exist, skipping") continue md = metadatas[inx] if metadatas else {} files: dict = { "file": (file, open(file, "rb")), "doc_metadata": json.dumps(md), } headers = self._get_post_headers() headers.pop("Content-Type") response = self._session.post( f"https://api.vectara.io/upload?c={self._vectara_customer_id}&o={self._vectara_corpus_id}&d=True", files=files, verify=True, headers=headers, timeout=self.vectara_api_timeout, ) if response.status_code == 409: doc_id = response.json()["document"]["documentId"] logger.info( f"File {file} already exists on Vectara (doc_id={doc_id}), skipping" ) elif response.status_code == 200: doc_id = response.json()["document"]["documentId"] doc_ids.append(doc_id) else: logger.info(f"Error indexing file {file}: {response.json()}") return doc_ids def add_texts( self, texts: Iterable[str], metadatas: Optional[List[dict]] = None, doc_metadata: Optional[dict] = None, **kwargs: Any, ) -> List[str]: """Run more texts through the embeddings and add to the vectorstore. Args: texts: Iterable of strings to add to the vectorstore. metadatas: Optional list of metadatas associated with the texts. doc_metadata: optional metadata for the document This function indexes all the input text strings in the Vectara corpus as a single Vectara document, where each input text is considered a "section" and the metadata are associated with each section. if 'doc_metadata' is provided, it is associated with the Vectara document. Returns: document ID of the document added """ doc_hash = md5() for t in texts: doc_hash.update(t.encode()) doc_id = doc_hash.hexdigest() if metadatas is None: metadatas = [{} for _ in texts] if doc_metadata: doc_metadata["source"] = "langchain" else: doc_metadata = {"source": "langchain"} use_core_api = kwargs.get("use_core_api", False) section_key = "parts" if use_core_api else "section" doc = { "document_id": doc_id, "metadataJson": json.dumps(doc_metadata), section_key: [ {"text": text, "metadataJson": json.dumps(md)} for text, md in zip(texts, metadatas) ], } success_str = self._index_doc(doc, use_core_api=use_core_api) if success_str == "E_ALREADY_EXISTS": self._delete_doc(doc_id) self._index_doc(doc) elif success_str == "E_NO_PERMISSIONS": print( # noqa: T201 """No permissions to add document to Vectara. Check your corpus ID, customer ID and API key""" ) return [doc_id] def _get_query_body( self, query: str, config: VectaraQueryConfig, chat: Optional[bool] = False, chat_conv_id: Optional[str] = None, **kwargs: Any, ) -> dict: """Build the body for the API Args: query: Text to look up documents similar to. config: VectaraQueryConfig object Returns: A dictionary with the body of the query """ if isinstance(config.rerank_config, dict): config.rerank_config = RerankConfig(**config.rerank_config) if isinstance(config.summary_config, dict): config.summary_config = SummaryConfig(**config.summary_config) body = { "query": [ { "query": query, "start": 0, "numResults": ( config.rerank_config.rerank_k if ( config.rerank_config.reranker in ["mmr", "udf", "rerank_multilingual_v1"] ) else config.k ), "contextConfig": { "sentencesBefore": config.n_sentence_before, "sentencesAfter": config.n_sentence_after, }, "corpusKey": [ { "corpusId": self._vectara_corpus_id, "metadataFilter": config.filter, } ], } ] } if config.lambda_val > 0: body["query"][0]["corpusKey"][0]["lexicalInterpolationConfig"] = { # type: ignore "lambda": config.lambda_val } if config.rerank_config.reranker == "mmr": body["query"][0]["rerankingConfig"] = { "rerankerId": MMR_RERANKER_ID, "mmrConfig": {"diversityBias": config.rerank_config.mmr_diversity_bias}, } elif config.rerank_config.reranker == "udf": body["query"][0]["rerankingConfig"] = { "rerankerId": UDF_RERANKER_ID, "userFunction": config.rerank_config.user_function, } elif config.rerank_config.reranker == "rerank_multilingual_v1": body["query"][0]["rerankingConfig"] = { "rerankerId": RERANKER_MULTILINGUAL_V1_ID, } if config.summary_config.is_enabled: body["query"][0]["summary"] = [ { "maxSummarizedResults": config.summary_config.max_results, "responseLang": config.summary_config.response_lang, "summarizerPromptName": config.summary_config.prompt_name, } ] if chat: body["query"][0]["summary"][0]["chat"] = { # type: ignore "store": True, "conversationId": chat_conv_id, } return body def vectara_query( self, query: str, config: VectaraQueryConfig, **kwargs: Any, ) -> List[Tuple[Document, float]]: """Run a Vectara query Args: query: Text to look up documents similar to. config: VectaraQueryConfig object Returns: A list of k Documents matching the given query If summary is enabled, last document is the summary text with 'summary'=True """ body = self._get_query_body(query, config, **kwargs) response = self._session.post( headers=self._get_post_headers(), url="https://api.vectara.io/v1/query", data=json.dumps(body), timeout=self.vectara_api_timeout, ) if response.status_code != 200: logger.error( "Query failed %s", f"(code {response.status_code}, reason {response.reason}, details " f"{response.text})", ) return [] result = response.json() if config.score_threshold: responses = [ r for r in result["responseSet"][0]["response"] if r["score"] > config.score_threshold ] else: responses = result["responseSet"][0]["response"] documents = result["responseSet"][0]["document"] metadatas = [] for x in responses: md = {m["name"]: m["value"] for m in x["metadata"]} doc_num = x["documentIndex"] doc_md = {m["name"]: m["value"] for m in documents[doc_num]["metadata"]} if "source" not in doc_md: doc_md["source"] = "vectara" md.update(doc_md) metadatas.append(md) res = [ ( Document( page_content=x["text"], metadata=md, ), x["score"], ) for x, md in zip(responses, metadatas) ] if config.rerank_config.reranker in ["mmr", "rerank_multilingual_v1"]: res = res[: config.k] if config.summary_config.is_enabled: summary = result["responseSet"][0]["summary"][0]["text"] fcs = result["responseSet"][0]["summary"][0]["factualConsistency"]["score"] res.append( ( Document( page_content=summary, metadata={"summary": True, "fcs": fcs} ), 0.0, ) ) return res def similarity_search_with_score( self, query: str, **kwargs: Any, ) -> List[Tuple[Document, float]]: """Return Vectara documents most similar to query, along with scores. Args: query: Text to look up documents similar to. k: Number of Documents to return. Defaults to 10. any other querying variable in VectaraQueryConfig like: - lambda_val: lexical match parameter for hybrid search. - filter: filter string - score_threshold: minimal score threshold for the result. - n_sentence_before: number of sentences before the matching segment - n_sentence_after: number of sentences after the matching segment - rerank_config: optional configuration for Reranking (see RerankConfig dataclass) - summary_config: optional configuration for summary (see SummaryConfig dataclass) Returns: List of Documents most similar to the query and score for each. """ config = VectaraQueryConfig(**kwargs) docs = self.vectara_query(query, config) return docs def similarity_search( # type: ignore[override] self, query: str, **kwargs: Any, ) -> List[Document]: """Return Vectara documents most similar to query, along with scores. Args: query: Text to look up documents similar to. any other querying variable in VectaraQueryConfig Returns: List of Documents most similar to the query """ docs_and_scores = self.similarity_search_with_score( query, **kwargs, ) return [doc for doc, _ in docs_and_scores] def max_marginal_relevance_search( # type: ignore[override] self, query: str, fetch_k: int = 50, lambda_mult: float = 0.5, **kwargs: Any, ) -> List[Document]: """Return docs selected using the maximal marginal relevance. Maximal marginal relevance optimizes for similarity to query AND diversity among selected documents. Args: query: Text to look up documents similar to. k: Number of Documents to return. Defaults to 5. fetch_k: Number of Documents to fetch to pass to MMR algorithm. Defaults to 50 lambda_mult: Number between 0 and 1 that determines the degree of diversity among the results with 0 corresponding to maximum diversity and 1 to minimum diversity. Defaults to 0.5. kwargs: any other querying variable in VectaraQueryConfig Returns: List of Documents selected by maximal marginal relevance. """ kwargs["rerank_config"] = RerankConfig( reranker="mmr", rerank_k=fetch_k, mmr_diversity_bias=1 - lambda_mult ) return self.similarity_search(query, **kwargs) @classmethod def from_texts( cls: Type[Vectara], texts: List[str], embedding: Optional[Embeddings] = None, metadatas: Optional[List[dict]] = None, **kwargs: Any, ) -> Vectara: """Construct Vectara wrapper from raw documents. This is intended to be a quick way to get started. Example: .. code-block:: python from langchain_community.vectorstores import Vectara vectara = Vectara.from_texts( texts, vectara_customer_id=customer_id, vectara_corpus_id=corpus_id, vectara_api_key=api_key, ) """ # Notes: # * Vectara generates its own embeddings, so we ignore the provided # embeddings (required by interface) # * when metadatas[] are provided they are associated with each "part" # in Vectara. doc_metadata can be used to provide additional metadata # for the document itself (applies to all "texts" in this call) doc_metadata = kwargs.pop("doc_metadata", {}) vectara = cls(**kwargs) vectara.add_texts(texts, metadatas, doc_metadata=doc_metadata, **kwargs) return vectara @classmethod def from_files( cls: Type[Vectara], files: List[str], embedding: Optional[Embeddings] = None, metadatas: Optional[List[dict]] = None, **kwargs: Any, ) -> Vectara: """Construct Vectara wrapper from raw documents. This is intended to be a quick way to get started. Example: .. code-block:: python from langchain_community.vectorstores import Vectara vectara = Vectara.from_files( files_list, vectara_customer_id=customer_id, vectara_corpus_id=corpus_id, vectara_api_key=api_key, ) """ # Note: Vectara generates its own embeddings, so we ignore the provided # embeddings (required by interface) vectara = cls(**kwargs) vectara.add_files(files, metadatas) return vectara def as_rag(self, config: VectaraQueryConfig) -> VectaraRAG: """Return a Vectara RAG runnable.""" return VectaraRAG(self, config) def as_chat(self, config: VectaraQueryConfig) -> VectaraRAG: """Return a Vectara RAG runnable for chat.""" return VectaraRAG(self, config, chat=True) def as_retriever(self, **kwargs: Any) -> VectaraRetriever: """return a retriever object.""" return VectaraRetriever( vectorstore=self, config=kwargs.get("config", VectaraQueryConfig()) ) class VectaraRetriever(VectorStoreRetriever): # type: ignore[override] """Vectara Retriever class.""" vectorstore: Vectara """VectorStore to use for retrieval.""" config: VectaraQueryConfig """Configuration for this retriever.""" model_config = ConfigDict( arbitrary_types_allowed=True, ) def _get_relevant_documents( self, query: str, *, run_manager: CallbackManagerForRetrieverRun ) -> List[Document]: docs_and_scores = self.vectorstore.vectara_query(query, self.config) return [doc for doc, _ in docs_and_scores] def add_documents(self, documents: List[Document], **kwargs: Any) -> List[str]: """Add documents to vectorstore.""" return self.vectorstore.add_documents(documents, **kwargs) class VectaraRAG(Runnable): """Vectara RAG runnable. Parameters: vectara: Vectara object config: VectaraQueryConfig object chat: bool, default False """ def __init__( self, vectara: Vectara, config: VectaraQueryConfig, chat: bool = False ): self.vectara = vectara self.config = config self.chat = chat self.conv_id = None def stream( self, input: str, config: Optional[RunnableConfig] = None, **kwargs: Any, ) -> Iterator[dict]: """Get streaming output from Vectara RAG. Args: input: The input query config: RunnableConfig object kwargs: Any additional arguments Returns: The output dictionary with question, answer and context """ body = self.vectara._get_query_body(input, self.config, self.chat, self.conv_id) response = self.vectara._session.post( headers=self.vectara._get_post_headers(), url="https://api.vectara.io/v1/stream-query", data=json.dumps(body), timeout=self.vectara.vectara_api_timeout, stream=True, ) if response.status_code != 200: logger.error( "Query failed %s", f"(code {response.status_code}, reason {response.reason}, details " f"{response.text})", ) return responses = [] documents = [] yield {"question": input} # First chunk is the question for line in response.iter_lines(): if line: # filter out keep-alive new lines data = json.loads(line.decode("utf-8")) result = data["result"] response_set = result["responseSet"] if response_set is None: summary = result.get("summary", None) if summary is None: continue if len(summary.get("status")) > 0: logger.error( f"Summary generation failed with status " f"{summary.get('status')[0].get('statusDetail')}" ) continue # Store conversation ID for chat, if applicable chat = summary.get("chat", None) if chat and chat.get("status", None): st_code = chat["status"] logger.info(f"Chat query failed with code {st_code}") if st_code == "RESOURCE_EXHAUSTED": self.conv_id = None logger.error( "Sorry, Vectara chat turns exceeds plan limit." ) continue conv_id = chat.get("conversationId", None) if chat else None if conv_id: self.conv_id = conv_id # If FCS is provided, pull it from the JSON response if summary.get("factualConsistency", None): fcs = summary.get("factualConsistency", {}).get("score", None) yield {"fcs": fcs} continue # Yield the summary chunk chunk = str(summary["text"]) yield {"answer": chunk} else: if self.config.score_threshold: responses = [ r for r in response_set["response"] if r["score"] > self.config.score_threshold ] else: responses = response_set["response"] documents = response_set["document"] metadatas = [] for x in responses: md = {m["name"]: m["value"] for m in x["metadata"]} doc_num = x["documentIndex"] doc_md = { m["name"]: m["value"] for m in documents[doc_num]["metadata"] } if "source" not in doc_md: doc_md["source"] = "vectara" md.update(doc_md) metadatas.append(md) res = [ ( Document( page_content=x["text"], metadata=md, ), x["score"], ) for x, md in zip(responses, metadatas) ] if self.config.rerank_config.reranker in [ "mmr", "rerank_multilingual_v1", ]: res = res[: self.config.k] yield {"context": res} return def invoke( self, input: str, config: Optional[RunnableConfig] = None, **kwargs: Any, ) -> dict: res = {"answer": ""} for chunk in self.stream(input): if "context" in chunk: res["context"] = chunk["context"] elif "question" in chunk: res["question"] = chunk["question"] elif "answer" in chunk: res["answer"] += chunk["answer"] elif "fcs" in chunk: res["fcs"] = chunk["fcs"] else: logger.error(f"Unknown chunk type: {chunk}") return res