diff --git a/docs/modules/indexes/retrievers/examples/contextual-compression.ipynb b/docs/modules/indexes/retrievers/examples/contextual-compression.ipynb new file mode 100644 index 00000000..9f299c6b --- /dev/null +++ b/docs/modules/indexes/retrievers/examples/contextual-compression.ipynb @@ -0,0 +1,371 @@ +{ + "cells": [ + { + "cell_type": "markdown", + "id": "fc0db1bc", + "metadata": {}, + "source": [ + "# Contextual Compression Retriever\n", + "\n", + "This notebook introduces the concept of DocumentCompressors and the ContextualCompressionRetriever. The core idea is simple: given a specific query, we should be able to return only the documents relevant to that query, and only the parts of those documents that are relevant. The ContextualCompressionsRetriever is a wrapper for another retriever that iterates over the initial output of the base retriever and filters and compresses those initial documents, so that only the most relevant information is returned." + ] + }, + { + "cell_type": "code", + "execution_count": 1, + "id": "28e8dc12", + "metadata": {}, + "outputs": [], + "source": [ + "# Helper function for printing docs\n", + "\n", + "def pretty_print_docs(docs):\n", + " print(f\"\\n{'-' * 100}\\n\".join([f\"Document {i+1}:\\n\\n\" + d.page_content for i, d in enumerate(docs)]))" + ] + }, + { + "cell_type": "markdown", + "id": "6fa3d916", + "metadata": {}, + "source": [ + "## Using a vanilla vector store retriever\n", + "Let's start by initializing a simple vector store retriever and storing the 2023 State of the Union speech (in chunks). We can see that given an example question our retriever returns one or two relevant docs and a few irrelevant docs. And even the relevant docs have a lot of irrelevant information in them." + ] + }, + { + "cell_type": "code", + "execution_count": 2, + "id": "9fbcc58f", + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Document 1:\n", + "\n", + "Tonight. I call on the Senate to: Pass the Freedom to Vote Act. Pass the John Lewis Voting Rights Act. And while you’re at it, pass the Disclose Act so Americans can know who is funding our elections. \n", + "\n", + "Tonight, I’d like to honor someone who has dedicated his life to serve this country: Justice Stephen Breyer—an Army veteran, Constitutional scholar, and retiring Justice of the United States Supreme Court. Justice Breyer, thank you for your service. \n", + "\n", + "One of the most serious constitutional responsibilities a President has is nominating someone to serve on the United States Supreme Court. \n", + "\n", + "And I did that 4 days ago, when I nominated Circuit Court of Appeals Judge Ketanji Brown Jackson. One of our nation’s top legal minds, who will continue Justice Breyer’s legacy of excellence.\n", + "----------------------------------------------------------------------------------------------------\n", + "Document 2:\n", + "\n", + "A former top litigator in private practice. A former federal public defender. And from a family of public school educators and police officers. A consensus builder. Since she’s been nominated, she’s received a broad range of support—from the Fraternal Order of Police to former judges appointed by Democrats and Republicans. \n", + "\n", + "And if we are to advance liberty and justice, we need to secure the Border and fix the immigration system. \n", + "\n", + "We can do both. At our border, we’ve installed new technology like cutting-edge scanners to better detect drug smuggling. \n", + "\n", + "We’ve set up joint patrols with Mexico and Guatemala to catch more human traffickers. \n", + "\n", + "We’re putting in place dedicated immigration judges so families fleeing persecution and violence can have their cases heard faster. \n", + "\n", + "We’re securing commitments and supporting partners in South and Central America to host more refugees and secure their own borders.\n", + "----------------------------------------------------------------------------------------------------\n", + "Document 3:\n", + "\n", + "And for our LGBTQ+ Americans, let’s finally get the bipartisan Equality Act to my desk. The onslaught of state laws targeting transgender Americans and their families is wrong. \n", + "\n", + "As I said last year, especially to our younger transgender Americans, I will always have your back as your President, so you can be yourself and reach your God-given potential. \n", + "\n", + "While it often appears that we never agree, that isn’t true. I signed 80 bipartisan bills into law last year. From preventing government shutdowns to protecting Asian-Americans from still-too-common hate crimes to reforming military justice. \n", + "\n", + "And soon, we’ll strengthen the Violence Against Women Act that I first wrote three decades ago. It is important for us to show the nation that we can come together and do big things. \n", + "\n", + "So tonight I’m offering a Unity Agenda for the Nation. Four big things we can do together. \n", + "\n", + "First, beat the opioid epidemic.\n", + "----------------------------------------------------------------------------------------------------\n", + "Document 4:\n", + "\n", + "Tonight, I’m announcing a crackdown on these companies overcharging American businesses and consumers. \n", + "\n", + "And as Wall Street firms take over more nursing homes, quality in those homes has gone down and costs have gone up. \n", + "\n", + "That ends on my watch. \n", + "\n", + "Medicare is going to set higher standards for nursing homes and make sure your loved ones get the care they deserve and expect. \n", + "\n", + "We’ll also cut costs and keep the economy going strong by giving workers a fair shot, provide more training and apprenticeships, hire them based on their skills not degrees. \n", + "\n", + "Let’s pass the Paycheck Fairness Act and paid leave. \n", + "\n", + "Raise the minimum wage to $15 an hour and extend the Child Tax Credit, so no one has to raise a family in poverty. \n", + "\n", + "Let’s increase Pell Grants and increase our historic support of HBCUs, and invest in what Jill—our First Lady who teaches full-time—calls America’s best-kept secret: community colleges.\n" + ] + } + ], + "source": [ + "from langchain.text_splitter import CharacterTextSplitter\n", + "from langchain.embeddings import OpenAIEmbeddings\n", + "from langchain.document_loaders import TextLoader\n", + "from langchain.vectorstores import FAISS\n", + "\n", + "documents = TextLoader('../../../state_of_the_union.txt').load()\n", + "text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0)\n", + "texts = text_splitter.split_documents(documents)\n", + "retriever = FAISS.from_documents(texts, OpenAIEmbeddings()).as_retriever()\n", + "\n", + "docs = retriever.get_relevant_documents(\"What did the president say about Ketanji Brown Jackson\")\n", + "pretty_print_docs(docs)" + ] + }, + { + "cell_type": "markdown", + "id": "b7648612", + "metadata": {}, + "source": [ + "## Adding contextual compression with an `LLMChainExtractor`\n", + "Now let's wrap our base retriever with a `ContextualCompressionRetriever`. We'll add an `LLMChainExtractor`, which will iterate over the initially returned documents and extract from each only the content that is relevant to the query." + ] + }, + { + "cell_type": "code", + "execution_count": 3, + "id": "9a658023", + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Document 1:\n", + "\n", + "\"One of the most serious constitutional responsibilities a President has is nominating someone to serve on the United States Supreme Court. \n", + "\n", + "And I did that 4 days ago, when I nominated Circuit Court of Appeals Judge Ketanji Brown Jackson. One of our nation’s top legal minds, who will continue Justice Breyer’s legacy of excellence.\"\n", + "----------------------------------------------------------------------------------------------------\n", + "Document 2:\n", + "\n", + "\"A former top litigator in private practice. A former federal public defender. And from a family of public school educators and police officers. A consensus builder. Since she’s been nominated, she’s received a broad range of support—from the Fraternal Order of Police to former judges appointed by Democrats and Republicans.\"\n" + ] + } + ], + "source": [ + "from langchain.llms import OpenAI\n", + "from langchain.retrievers import ContextualCompressionRetriever\n", + "from langchain.retrievers.document_compressors import LLMChainExtractor\n", + "\n", + "llm = OpenAI(temperature=0)\n", + "compressor = LLMChainExtractor.from_llm(llm)\n", + "compression_retriever = ContextualCompressionRetriever(base_compressor=compressor, base_retriever=retriever)\n", + "\n", + "compressed_docs = compression_retriever.get_relevant_documents(\"What did the president say about Ketanji Jackson Brown\")\n", + "pretty_print_docs(compressed_docs)" + ] + }, + { + "cell_type": "markdown", + "id": "2cd38f3a", + "metadata": {}, + "source": [ + "## More built-in compressors: filters\n", + "### `LLMChainFilter`\n", + "The `LLMChainFilter` is slightly simpler but more robust compressor that uses an LLM chain to decide which of the initially retrieved documents to filter out and which ones to return, without manipulating the document contents." + ] + }, + { + "cell_type": "code", + "execution_count": 4, + "id": "b216a767", + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Document 1:\n", + "\n", + "Tonight. I call on the Senate to: Pass the Freedom to Vote Act. Pass the John Lewis Voting Rights Act. And while you’re at it, pass the Disclose Act so Americans can know who is funding our elections. \n", + "\n", + "Tonight, I’d like to honor someone who has dedicated his life to serve this country: Justice Stephen Breyer—an Army veteran, Constitutional scholar, and retiring Justice of the United States Supreme Court. Justice Breyer, thank you for your service. \n", + "\n", + "One of the most serious constitutional responsibilities a President has is nominating someone to serve on the United States Supreme Court. \n", + "\n", + "And I did that 4 days ago, when I nominated Circuit Court of Appeals Judge Ketanji Brown Jackson. One of our nation’s top legal minds, who will continue Justice Breyer’s legacy of excellence.\n" + ] + } + ], + "source": [ + "from langchain.retrievers.document_compressors import LLMChainFilter\n", + "\n", + "_filter = LLMChainFilter.from_llm(llm)\n", + "compression_retriever = ContextualCompressionRetriever(base_compressor=_filter, base_retriever=retriever)\n", + "\n", + "compressed_docs = compression_retriever.get_relevant_documents(\"What did the president say about Ketanji Jackson Brown\")\n", + "pretty_print_docs(compressed_docs)" + ] + }, + { + "cell_type": "markdown", + "id": "8c709598", + "metadata": {}, + "source": [ + "### `EmbeddingsFilter`\n", + "\n", + "Making an extra LLM call over each retrieved document is expensive and slow. The `EmbeddingsFilter` provides a cheaper and faster option by embedding the documents and query and only returning those documents which have sufficiently similar embeddings to the query." + ] + }, + { + "cell_type": "code", + "execution_count": 5, + "id": "6fbc801f", + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Document 1:\n", + "\n", + "Tonight. I call on the Senate to: Pass the Freedom to Vote Act. Pass the John Lewis Voting Rights Act. And while you’re at it, pass the Disclose Act so Americans can know who is funding our elections. \n", + "\n", + "Tonight, I’d like to honor someone who has dedicated his life to serve this country: Justice Stephen Breyer—an Army veteran, Constitutional scholar, and retiring Justice of the United States Supreme Court. Justice Breyer, thank you for your service. \n", + "\n", + "One of the most serious constitutional responsibilities a President has is nominating someone to serve on the United States Supreme Court. \n", + "\n", + "And I did that 4 days ago, when I nominated Circuit Court of Appeals Judge Ketanji Brown Jackson. One of our nation’s top legal minds, who will continue Justice Breyer’s legacy of excellence.\n", + "----------------------------------------------------------------------------------------------------\n", + "Document 2:\n", + "\n", + "A former top litigator in private practice. A former federal public defender. And from a family of public school educators and police officers. A consensus builder. Since she’s been nominated, she’s received a broad range of support—from the Fraternal Order of Police to former judges appointed by Democrats and Republicans. \n", + "\n", + "And if we are to advance liberty and justice, we need to secure the Border and fix the immigration system. \n", + "\n", + "We can do both. At our border, we’ve installed new technology like cutting-edge scanners to better detect drug smuggling. \n", + "\n", + "We’ve set up joint patrols with Mexico and Guatemala to catch more human traffickers. \n", + "\n", + "We’re putting in place dedicated immigration judges so families fleeing persecution and violence can have their cases heard faster. \n", + "\n", + "We’re securing commitments and supporting partners in South and Central America to host more refugees and secure their own borders.\n", + "----------------------------------------------------------------------------------------------------\n", + "Document 3:\n", + "\n", + "And for our LGBTQ+ Americans, let’s finally get the bipartisan Equality Act to my desk. The onslaught of state laws targeting transgender Americans and their families is wrong. \n", + "\n", + "As I said last year, especially to our younger transgender Americans, I will always have your back as your President, so you can be yourself and reach your God-given potential. \n", + "\n", + "While it often appears that we never agree, that isn’t true. I signed 80 bipartisan bills into law last year. From preventing government shutdowns to protecting Asian-Americans from still-too-common hate crimes to reforming military justice. \n", + "\n", + "And soon, we’ll strengthen the Violence Against Women Act that I first wrote three decades ago. It is important for us to show the nation that we can come together and do big things. \n", + "\n", + "So tonight I’m offering a Unity Agenda for the Nation. Four big things we can do together. \n", + "\n", + "First, beat the opioid epidemic.\n" + ] + } + ], + "source": [ + "from langchain.embeddings import OpenAIEmbeddings\n", + "from langchain.retrievers.document_compressors import EmbeddingsFilter\n", + "\n", + "embeddings = OpenAIEmbeddings()\n", + "embeddings_filter = EmbeddingsFilter(embeddings=embeddings, similarity_threshold=0.76)\n", + "compression_retriever = ContextualCompressionRetriever(base_compressor=embeddings_filter, base_retriever=retriever)\n", + "\n", + "compressed_docs = compression_retriever.get_relevant_documents(\"What did the president say about Ketanji Jackson Brown\")\n", + "pretty_print_docs(compressed_docs)" + ] + }, + { + "cell_type": "markdown", + "id": "07365d36", + "metadata": {}, + "source": [ + "# Stringing compressors and document transformers together\n", + "Using the `DocumentCompressorPipeline` we can also easily combine multiple compressors in sequence. Along with compressors we can add `BaseDocumentTransformer`s to our pipeline, which don't perform any contextual compression but simply perform some transformation on a set of documents. For example `TextSplitter`s can be used as document transformers to split documents into smaller pieces, and the `EmbeddingsRedundantFilter` can be used to filter out redundant documents based on embedding similarity between documents.\n", + "\n", + "Below we create a compressor pipeline by first splitting our docs into smaller chunks, then removing redundant documents, and then filtering based on relevance to the query." + ] + }, + { + "cell_type": "code", + "execution_count": 6, + "id": "2a150a63", + "metadata": {}, + "outputs": [], + "source": [ + "from langchain.document_transformers import EmbeddingsRedundantFilter\n", + "from langchain.retrievers.document_compressors import DocumentCompressorPipeline\n", + "from langchain.text_splitter import CharacterTextSplitter\n", + "\n", + "splitter = CharacterTextSplitter(chunk_size=300, chunk_overlap=0, separator=\". \")\n", + "redundant_filter = EmbeddingsRedundantFilter(embeddings=embeddings)\n", + "relevant_filter = EmbeddingsFilter(embeddings=embeddings, similarity_threshold=0.76)\n", + "pipeline_compressor = DocumentCompressorPipeline(\n", + " transformers=[splitter, redundant_filter, relevant_filter]\n", + ")" + ] + }, + { + "cell_type": "code", + "execution_count": 7, + "id": "3ceab64a", + "metadata": {}, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Document 1:\n", + "\n", + "One of the most serious constitutional responsibilities a President has is nominating someone to serve on the United States Supreme Court. \n", + "\n", + "And I did that 4 days ago, when I nominated Circuit Court of Appeals Judge Ketanji Brown Jackson\n", + "----------------------------------------------------------------------------------------------------\n", + "Document 2:\n", + "\n", + "As I said last year, especially to our younger transgender Americans, I will always have your back as your President, so you can be yourself and reach your God-given potential. \n", + "\n", + "While it often appears that we never agree, that isn’t true. I signed 80 bipartisan bills into law last year\n", + "----------------------------------------------------------------------------------------------------\n", + "Document 3:\n", + "\n", + "A former top litigator in private practice. A former federal public defender. And from a family of public school educators and police officers. A consensus builder\n" + ] + } + ], + "source": [ + "compression_retriever = ContextualCompressionRetriever(base_compressor=pipeline_compressor, base_retriever=retriever)\n", + "\n", + "compressed_docs = compression_retriever.get_relevant_documents(\"What did the president say about Ketanji Jackson Brown\")\n", + "pretty_print_docs(compressed_docs)" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "8cfd9fc5", + "metadata": {}, + "outputs": [], + "source": [] + } + ], + "metadata": { + "kernelspec": { + "display_name": "Python 3 (ipykernel)", + "language": "python", + "name": "python3" + }, + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.9.1" + } + }, + "nbformat": 4, + "nbformat_minor": 5 +} diff --git a/langchain/document_transformers.py b/langchain/document_transformers.py new file mode 100644 index 00000000..7f17cb68 --- /dev/null +++ b/langchain/document_transformers.py @@ -0,0 +1,100 @@ +"""Transform documents""" +from typing import Any, Callable, List, Sequence + +import numpy as np +from pydantic import BaseModel, Field + +from langchain.embeddings.base import Embeddings +from langchain.math_utils import cosine_similarity +from langchain.schema import BaseDocumentTransformer, Document + + +class _DocumentWithState(Document): + """Wrapper for a document that includes arbitrary state.""" + + state: dict = Field(default_factory=dict) + """State associated with the document.""" + + def to_document(self) -> Document: + """Convert the DocumentWithState to a Document.""" + return Document(page_content=self.page_content, metadata=self.metadata) + + @classmethod + def from_document(cls, doc: Document) -> "_DocumentWithState": + """Create a DocumentWithState from a Document.""" + if isinstance(doc, cls): + return doc + return cls(page_content=doc.page_content, metadata=doc.metadata) + + +def get_stateful_documents( + documents: Sequence[Document], +) -> Sequence[_DocumentWithState]: + return [_DocumentWithState.from_document(doc) for doc in documents] + + +def _filter_similar_embeddings( + embedded_documents: List[List[float]], similarity_fn: Callable, threshold: float +) -> List[int]: + """Filter redundant documents based on the similarity of their embeddings.""" + similarity = np.tril(similarity_fn(embedded_documents, embedded_documents), k=-1) + redundant = np.where(similarity > threshold) + redundant_stacked = np.column_stack(redundant) + redundant_sorted = np.argsort(similarity[redundant])[::-1] + included_idxs = set(range(len(embedded_documents))) + for first_idx, second_idx in redundant_stacked[redundant_sorted]: + if first_idx in included_idxs and second_idx in included_idxs: + # Default to dropping the second document of any highly similar pair. + included_idxs.remove(second_idx) + return list(sorted(included_idxs)) + + +def _get_embeddings_from_stateful_docs( + embeddings: Embeddings, documents: Sequence[_DocumentWithState] +) -> List[List[float]]: + if len(documents) and "embedded_doc" in documents[0].state: + embedded_documents = [doc.state["embedded_doc"] for doc in documents] + else: + embedded_documents = embeddings.embed_documents( + [d.page_content for d in documents] + ) + for doc, embedding in zip(documents, embedded_documents): + doc.state["embedded_doc"] = embedding + return embedded_documents + + +class EmbeddingsRedundantFilter(BaseDocumentTransformer, BaseModel): + """Filter that drops redundant documents by comparing their embeddings.""" + + embeddings: Embeddings + """Embeddings to use for embedding document contents.""" + similarity_fn: Callable = cosine_similarity + """Similarity function for comparing documents. Function expected to take as input + two matrices (List[List[float]]) and return a matrix of scores where higher values + indicate greater similarity.""" + similarity_threshold: float = 0.95 + """Threshold for determining when two documents are similar enough + to be considered redundant.""" + + class Config: + """Configuration for this pydantic object.""" + + arbitrary_types_allowed = True + + def transform_documents( + self, documents: Sequence[Document], **kwargs: Any + ) -> Sequence[Document]: + """Filter down documents.""" + stateful_documents = get_stateful_documents(documents) + embedded_documents = _get_embeddings_from_stateful_docs( + self.embeddings, stateful_documents + ) + included_idxs = _filter_similar_embeddings( + embedded_documents, self.similarity_fn, self.similarity_threshold + ) + return [stateful_documents[i] for i in sorted(included_idxs)] + + async def atransform_documents( + self, documents: Sequence[Document], **kwargs: Any + ) -> Sequence[Document]: + raise NotImplementedError diff --git a/langchain/math_utils.py b/langchain/math_utils.py new file mode 100644 index 00000000..218af047 --- /dev/null +++ b/langchain/math_utils.py @@ -0,0 +1,22 @@ +"""Math utils.""" +from typing import List, Union + +import numpy as np + +Matrix = Union[List[List[float]], List[np.ndarray], np.ndarray] + + +def cosine_similarity(X: Matrix, Y: Matrix) -> np.ndarray: + """Row-wise cosine similarity between two equal-width matrices.""" + if len(X) == 0 or len(Y) == 0: + return np.array([]) + X = np.array(X) + Y = np.array(Y) + if X.shape[1] != Y.shape[1]: + raise ValueError("Number of columns in X and Y must be the same.") + + X_norm = np.linalg.norm(X, axis=1) + Y_norm = np.linalg.norm(Y, axis=1) + similarity = np.dot(X, Y.T) / np.outer(X_norm, Y_norm) + similarity[np.isnan(similarity) | np.isinf(similarity)] = 0.0 + return similarity diff --git a/langchain/output_parsers/boolean.py b/langchain/output_parsers/boolean.py new file mode 100644 index 00000000..40890a9d --- /dev/null +++ b/langchain/output_parsers/boolean.py @@ -0,0 +1,29 @@ +from langchain.schema import BaseOutputParser + + +class BooleanOutputParser(BaseOutputParser[bool]): + true_val: str = "YES" + false_val: str = "NO" + + def parse(self, text: str) -> bool: + """Parse the output of an LLM call to a boolean. + + Args: + text: output of language model + + Returns: + boolean + + """ + cleaned_text = text.strip() + if cleaned_text not in (self.true_val, self.false_val): + raise ValueError( + f"BooleanOutputParser expected output value to either be " + f"{self.true_val} or {self.false_val}. Received {cleaned_text}." + ) + return cleaned_text == self.true_val + + @property + def _type(self) -> str: + """Snake-case string identifier for output parser type.""" + return "boolean_output_parser" diff --git a/langchain/retrievers/__init__.py b/langchain/retrievers/__init__.py index 869ea937..d89cf9d8 100644 --- a/langchain/retrievers/__init__.py +++ b/langchain/retrievers/__init__.py @@ -1,4 +1,5 @@ from langchain.retrievers.chatgpt_plugin_retriever import ChatGPTPluginRetriever +from langchain.retrievers.contextual_compression import ContextualCompressionRetriever from langchain.retrievers.databerry import DataberryRetriever from langchain.retrievers.elastic_search_bm25 import ElasticSearchBM25Retriever from langchain.retrievers.metal import MetalRetriever @@ -13,6 +14,7 @@ from langchain.retrievers.weaviate_hybrid_search import WeaviateHybridSearchRetr __all__ = [ "ChatGPTPluginRetriever", + "ContextualCompressionRetriever", "RemoteLangChainRetriever", "PineconeHybridSearchRetriever", "MetalRetriever", diff --git a/langchain/retrievers/contextual_compression.py b/langchain/retrievers/contextual_compression.py new file mode 100644 index 00000000..788a3919 --- /dev/null +++ b/langchain/retrievers/contextual_compression.py @@ -0,0 +1,51 @@ +"""Retriever that wraps a base retriever and filters the results.""" +from typing import List + +from pydantic import BaseModel, Extra + +from langchain.retrievers.document_compressors.base import ( + BaseDocumentCompressor, +) +from langchain.schema import BaseRetriever, Document + + +class ContextualCompressionRetriever(BaseRetriever, BaseModel): + """Retriever that wraps a base retriever and compresses the results.""" + + base_compressor: BaseDocumentCompressor + """Compressor for compressing retrieved documents.""" + + base_retriever: BaseRetriever + """Base Retriever to use for getting relevant documents.""" + + class Config: + """Configuration for this pydantic object.""" + + extra = Extra.forbid + arbitrary_types_allowed = True + + def get_relevant_documents(self, query: str) -> List[Document]: + """Get documents relevant for a query. + + Args: + query: string to find relevant documents for + + Returns: + Sequence of relevant documents + """ + docs = self.base_retriever.get_relevant_documents(query) + compressed_docs = self.base_compressor.compress_documents(docs, query) + return list(compressed_docs) + + async def aget_relevant_documents(self, query: str) -> List[Document]: + """Get documents relevant for a query. + + Args: + query: string to find relevant documents for + + Returns: + List of relevant documents + """ + docs = await self.base_retriever.aget_relevant_documents(query) + compressed_docs = await self.base_compressor.acompress_documents(docs, query) + return list(compressed_docs) diff --git a/langchain/retrievers/document_compressors/__init__.py b/langchain/retrievers/document_compressors/__init__.py new file mode 100644 index 00000000..528eae71 --- /dev/null +++ b/langchain/retrievers/document_compressors/__init__.py @@ -0,0 +1,17 @@ +from langchain.retrievers.document_compressors.base import DocumentCompressorPipeline +from langchain.retrievers.document_compressors.chain_extract import ( + LLMChainExtractor, +) +from langchain.retrievers.document_compressors.chain_filter import ( + LLMChainFilter, +) +from langchain.retrievers.document_compressors.embeddings_filter import ( + EmbeddingsFilter, +) + +__all__ = [ + "DocumentCompressorPipeline", + "EmbeddingsFilter", + "LLMChainExtractor", + "LLMChainFilter", +] diff --git a/langchain/retrievers/document_compressors/base.py b/langchain/retrievers/document_compressors/base.py new file mode 100644 index 00000000..b42d95ea --- /dev/null +++ b/langchain/retrievers/document_compressors/base.py @@ -0,0 +1,61 @@ +"""Interface for retrieved document compressors.""" +from abc import ABC, abstractmethod +from typing import List, Sequence, Union + +from pydantic import BaseModel + +from langchain.schema import BaseDocumentTransformer, Document + + +class BaseDocumentCompressor(BaseModel, ABC): + """""" + + @abstractmethod + def compress_documents( + self, documents: Sequence[Document], query: str + ) -> Sequence[Document]: + """Compress retrieved documents given the query context.""" + + @abstractmethod + async def acompress_documents( + self, documents: Sequence[Document], query: str + ) -> Sequence[Document]: + """Compress retrieved documents given the query context.""" + + +class DocumentCompressorPipeline(BaseDocumentCompressor): + """Document compressor that uses a pipeline of transformers.""" + + transformers: List[Union[BaseDocumentTransformer, BaseDocumentCompressor]] + """List of document filters that are chained together and run in sequence.""" + + class Config: + """Configuration for this pydantic object.""" + + arbitrary_types_allowed = True + + def compress_documents( + self, documents: Sequence[Document], query: str + ) -> Sequence[Document]: + """Transform a list of documents.""" + for _transformer in self.transformers: + if isinstance(_transformer, BaseDocumentCompressor): + documents = _transformer.compress_documents(documents, query) + elif isinstance(_transformer, BaseDocumentTransformer): + documents = _transformer.transform_documents(documents) + else: + raise ValueError(f"Got unexpected transformer type: {_transformer}") + return documents + + async def acompress_documents( + self, documents: Sequence[Document], query: str + ) -> Sequence[Document]: + """Compress retrieved documents given the query context.""" + for _transformer in self.transformers: + if isinstance(_transformer, BaseDocumentCompressor): + documents = await _transformer.acompress_documents(documents, query) + elif isinstance(_transformer, BaseDocumentTransformer): + documents = await _transformer.atransform_documents(documents) + else: + raise ValueError(f"Got unexpected transformer type: {_transformer}") + return documents diff --git a/langchain/retrievers/document_compressors/chain_extract.py b/langchain/retrievers/document_compressors/chain_extract.py new file mode 100644 index 00000000..6f638559 --- /dev/null +++ b/langchain/retrievers/document_compressors/chain_extract.py @@ -0,0 +1,77 @@ +"""DocumentFilter that uses an LLM chain to extract the relevant parts of documents.""" +from typing import Any, Callable, Dict, Optional, Sequence + +from langchain import LLMChain, PromptTemplate +from langchain.retrievers.document_compressors.base import ( + BaseDocumentCompressor, +) +from langchain.retrievers.document_compressors.chain_extract_prompt import ( + prompt_template, +) +from langchain.schema import BaseLanguageModel, BaseOutputParser, Document + + +def default_get_input(query: str, doc: Document) -> Dict[str, Any]: + """Return the compression chain input.""" + return {"question": query, "context": doc.page_content} + + +class NoOutputParser(BaseOutputParser[str]): + """Parse outputs that could return a null string of some sort.""" + + no_output_str: str = "NO_OUTPUT" + + def parse(self, text: str) -> str: + cleaned_text = text.strip() + if cleaned_text == self.no_output_str: + return "" + return cleaned_text + + +def _get_default_chain_prompt() -> PromptTemplate: + output_parser = NoOutputParser() + template = prompt_template.format(no_output_str=output_parser.no_output_str) + return PromptTemplate( + template=template, + input_variables=["question", "context"], + output_parser=output_parser, + ) + + +class LLMChainExtractor(BaseDocumentCompressor): + llm_chain: LLMChain + """LLM wrapper to use for compressing documents.""" + + get_input: Callable[[str, Document], dict] = default_get_input + """Callable for constructing the chain input from the query and a Document.""" + + def compress_documents( + self, documents: Sequence[Document], query: str + ) -> Sequence[Document]: + """Compress page content of raw documents.""" + compressed_docs = [] + for doc in documents: + _input = self.get_input(query, doc) + output = self.llm_chain.predict_and_parse(**_input) + if len(output) == 0: + continue + compressed_docs.append(Document(page_content=output, metadata=doc.metadata)) + return compressed_docs + + async def acompress_documents( + self, documents: Sequence[Document], query: str + ) -> Sequence[Document]: + raise NotImplementedError + + @classmethod + def from_llm( + cls, + llm: BaseLanguageModel, + prompt: Optional[PromptTemplate] = None, + get_input: Optional[Callable[[str, Document], str]] = None, + ) -> "LLMChainExtractor": + """Initialize from LLM.""" + _prompt = prompt if prompt is not None else _get_default_chain_prompt() + _get_input = get_input if get_input is not None else default_get_input + llm_chain = LLMChain(llm=llm, prompt=_prompt) + return cls(llm_chain=llm_chain, get_input=_get_input) diff --git a/langchain/retrievers/document_compressors/chain_extract_prompt.py b/langchain/retrievers/document_compressors/chain_extract_prompt.py new file mode 100644 index 00000000..c27b8770 --- /dev/null +++ b/langchain/retrievers/document_compressors/chain_extract_prompt.py @@ -0,0 +1,11 @@ +# flake8: noqa +prompt_template = """Given the following question and context, extract any part of the context *AS IS* that is relevant to answer the question. If none of the context is relevant return {no_output_str}. + +Remember, *DO NOT* edit the extracted parts of the context. + +> Question: {{question}} +> Context: +>>> +{{context}} +>>> +Extracted relevant parts:""" diff --git a/langchain/retrievers/document_compressors/chain_filter.py b/langchain/retrievers/document_compressors/chain_filter.py new file mode 100644 index 00000000..f5e33e6b --- /dev/null +++ b/langchain/retrievers/document_compressors/chain_filter.py @@ -0,0 +1,65 @@ +"""Filter that uses an LLM to drop documents that aren't relevant to the query.""" +from typing import Any, Callable, Dict, Optional, Sequence + +from langchain import BasePromptTemplate, LLMChain, PromptTemplate +from langchain.output_parsers.boolean import BooleanOutputParser +from langchain.retrievers.document_compressors.base import ( + BaseDocumentCompressor, +) +from langchain.retrievers.document_compressors.chain_filter_prompt import ( + prompt_template, +) +from langchain.schema import BaseLanguageModel, Document + + +def _get_default_chain_prompt() -> PromptTemplate: + return PromptTemplate( + template=prompt_template, + input_variables=["question", "context"], + output_parser=BooleanOutputParser(), + ) + + +def default_get_input(query: str, doc: Document) -> Dict[str, Any]: + """Return the compression chain input.""" + return {"question": query, "context": doc.page_content} + + +class LLMChainFilter(BaseDocumentCompressor): + """Filter that drops documents that aren't relevant to the query.""" + + llm_chain: LLMChain + """LLM wrapper to use for filtering documents. + The chain prompt is expected to have a BooleanOutputParser.""" + + get_input: Callable[[str, Document], dict] = default_get_input + """Callable for constructing the chain input from the query and a Document.""" + + def compress_documents( + self, documents: Sequence[Document], query: str + ) -> Sequence[Document]: + """Filter down documents based on their relevance to the query.""" + filtered_docs = [] + for doc in documents: + _input = self.get_input(query, doc) + include_doc = self.llm_chain.predict_and_parse(**_input) + if include_doc: + filtered_docs.append(doc) + return filtered_docs + + async def acompress_documents( + self, documents: Sequence[Document], query: str + ) -> Sequence[Document]: + """Filter down documents.""" + raise NotImplementedError + + @classmethod + def from_llm( + cls, + llm: BaseLanguageModel, + prompt: Optional[BasePromptTemplate] = None, + **kwargs: Any + ) -> "LLMChainFilter": + _prompt = prompt if prompt is not None else _get_default_chain_prompt() + llm_chain = LLMChain(llm=llm, prompt=_prompt) + return cls(llm_chain=llm_chain, **kwargs) diff --git a/langchain/retrievers/document_compressors/chain_filter_prompt.py b/langchain/retrievers/document_compressors/chain_filter_prompt.py new file mode 100644 index 00000000..5376dfa2 --- /dev/null +++ b/langchain/retrievers/document_compressors/chain_filter_prompt.py @@ -0,0 +1,9 @@ +# flake8: noqa +prompt_template = """Given the following question and context, return YES if the context is relevant to the question and NO if it isn't. + +> Question: {question} +> Context: +>>> +{context} +>>> +> Relevant (YES / NO):""" diff --git a/langchain/retrievers/document_compressors/embeddings_filter.py b/langchain/retrievers/document_compressors/embeddings_filter.py new file mode 100644 index 00000000..54338018 --- /dev/null +++ b/langchain/retrievers/document_compressors/embeddings_filter.py @@ -0,0 +1,70 @@ +"""Document compressor that uses embeddings to drop documents unrelated to the query.""" +from typing import Callable, Dict, Optional, Sequence + +import numpy as np +from pydantic import root_validator + +from langchain.document_transformers import ( + _get_embeddings_from_stateful_docs, + get_stateful_documents, +) +from langchain.embeddings.base import Embeddings +from langchain.math_utils import cosine_similarity +from langchain.retrievers.document_compressors.base import ( + BaseDocumentCompressor, +) +from langchain.schema import Document + + +class EmbeddingsFilter(BaseDocumentCompressor): + embeddings: Embeddings + """Embeddings to use for embedding document contents and queries.""" + similarity_fn: Callable = cosine_similarity + """Similarity function for comparing documents. Function expected to take as input + two matrices (List[List[float]]) and return a matrix of scores where higher values + indicate greater similarity.""" + k: Optional[int] = 20 + """The number of relevant documents to return. Can be set to None, in which case + `similarity_threshold` must be specified. Defaults to 20.""" + similarity_threshold: Optional[float] + """Threshold for determining when two documents are similar enough + to be considered redundant. Defaults to None, must be specified if `k` is set + to None.""" + + class Config: + """Configuration for this pydantic object.""" + + arbitrary_types_allowed = True + + @root_validator() + def validate_params(cls, values: Dict) -> Dict: + """Validate similarity parameters.""" + if values["k"] is None and values["similarity_threshold"] is None: + raise ValueError("Must specify one of `k` or `similarity_threshold`.") + return values + + def compress_documents( + self, documents: Sequence[Document], query: str + ) -> Sequence[Document]: + """Filter documents based on similarity of their embeddings to the query.""" + stateful_documents = get_stateful_documents(documents) + embedded_documents = _get_embeddings_from_stateful_docs( + self.embeddings, stateful_documents + ) + embedded_query = self.embeddings.embed_query(query) + similarity = self.similarity_fn([embedded_query], embedded_documents)[0] + included_idxs = np.arange(len(embedded_documents)) + if self.k is not None: + included_idxs = np.argsort(similarity)[::-1][: self.k] + if self.similarity_threshold is not None: + similar_enough = np.where( + similarity[included_idxs] > self.similarity_threshold + ) + included_idxs = included_idxs[similar_enough] + return [stateful_documents[i] for i in included_idxs] + + async def acompress_documents( + self, documents: Sequence[Document], query: str + ) -> Sequence[Document]: + """Filter down documents.""" + raise NotImplementedError diff --git a/langchain/schema.py b/langchain/schema.py index 65f53094..821dc70a 100644 --- a/langchain/schema.py +++ b/langchain/schema.py @@ -2,7 +2,17 @@ from __future__ import annotations from abc import ABC, abstractmethod -from typing import Any, Dict, Generic, List, NamedTuple, Optional, TypeVar, Union +from typing import ( + Any, + Dict, + Generic, + List, + NamedTuple, + Optional, + Sequence, + TypeVar, + Union, +) from pydantic import BaseModel, Extra, Field, root_validator @@ -394,16 +404,17 @@ class OutputParserException(Exception): pass -D = TypeVar("D", bound=Document) - - -class BaseDocumentTransformer(ABC, Generic[D]): +class BaseDocumentTransformer(ABC): """Base interface for transforming documents.""" @abstractmethod - def transform_documents(self, documents: List[D], **kwargs: Any) -> List[D]: + def transform_documents( + self, documents: Sequence[Document], **kwargs: Any + ) -> Sequence[Document]: """Transform a list of documents.""" @abstractmethod - async def atransform_documents(self, documents: List[D], **kwargs: Any) -> List[D]: + async def atransform_documents( + self, documents: Sequence[Document], **kwargs: Any + ) -> Sequence[Document]: """Asynchronously transform a list of documents.""" diff --git a/langchain/text_splitter.py b/langchain/text_splitter.py index b9c3c24b..7afd8b25 100644 --- a/langchain/text_splitter.py +++ b/langchain/text_splitter.py @@ -13,6 +13,7 @@ from typing import ( List, Literal, Optional, + Sequence, Union, ) @@ -22,7 +23,7 @@ from langchain.schema import BaseDocumentTransformer logger = logging.getLogger(__name__) -class TextSplitter(BaseDocumentTransformer[Document], ABC): +class TextSplitter(BaseDocumentTransformer, ABC): """Interface for splitting text into chunks.""" def __init__( @@ -63,7 +64,7 @@ class TextSplitter(BaseDocumentTransformer[Document], ABC): """Split documents.""" texts = [doc.page_content for doc in documents] metadatas = [doc.metadata for doc in documents] - return self.create_documents(texts, metadatas) + return self.create_documents(texts, metadatas=metadatas) def _join_docs(self, docs: List[str], separator: str) -> Optional[str]: text = separator.join(docs) @@ -173,15 +174,15 @@ class TextSplitter(BaseDocumentTransformer[Document], ABC): return cls(length_function=_tiktoken_encoder, **kwargs) def transform_documents( - self, documents: List[Document], **kwargs: Any - ) -> List[Document]: - """Transform list of documents by splitting them.""" - return self.split_documents(documents) + self, documents: Sequence[Document], **kwargs: Any + ) -> Sequence[Document]: + """Transform sequence of documents by splitting them.""" + return self.split_documents(list(documents)) async def atransform_documents( - self, documents: List[Document], **kwargs: Any - ) -> List[Document]: - """Asynchronously transform a list of documents by splitting them.""" + self, documents: Sequence[Document], **kwargs: Any + ) -> Sequence[Document]: + """Asynchronously transform a sequence of documents by splitting them.""" raise NotImplementedError diff --git a/langchain/vectorstores/utils.py b/langchain/vectorstores/utils.py index e34a7703..50e8ae6c 100644 --- a/langchain/vectorstores/utils.py +++ b/langchain/vectorstores/utils.py @@ -4,10 +4,7 @@ from typing import List import numpy as np - -def cosine_similarity(a: np.ndarray, b: np.ndarray) -> float: - """Calculate cosine similarity with numpy.""" - return np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b)) +from langchain.math_utils import cosine_similarity def maximal_marginal_relevance( @@ -17,22 +14,26 @@ def maximal_marginal_relevance( k: int = 4, ) -> List[int]: """Calculate maximal marginal relevance.""" - idxs: List[int] = [] - while len(idxs) < k: + if min(k, len(embedding_list)) <= 0: + return [] + similarity_to_query = cosine_similarity([query_embedding], embedding_list)[0] + most_similar = int(np.argmax(similarity_to_query)) + idxs = [most_similar] + selected = np.array([embedding_list[most_similar]]) + while len(idxs) < min(k, len(embedding_list)): best_score = -np.inf idx_to_add = -1 - for i, emb in enumerate(embedding_list): + similarity_to_selected = cosine_similarity(embedding_list, selected) + for i, query_score in enumerate(similarity_to_query): if i in idxs: continue - first_part = cosine_similarity(query_embedding, emb) - second_part = 0.0 - for j in idxs: - cos_sim = cosine_similarity(emb, embedding_list[j]) - if cos_sim > second_part: - second_part = cos_sim - equation_score = lambda_mult * first_part - (1 - lambda_mult) * second_part + redundant_score = max(similarity_to_selected[i]) + equation_score = ( + lambda_mult * query_score - (1 - lambda_mult) * redundant_score + ) if equation_score > best_score: best_score = equation_score idx_to_add = i idxs.append(idx_to_add) + selected = np.append(selected, [embedding_list[idx_to_add]], axis=0) return idxs diff --git a/tests/integration_tests/retrievers/__init__.py b/tests/integration_tests/retrievers/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/tests/integration_tests/retrievers/document_compressors/__init__.py b/tests/integration_tests/retrievers/document_compressors/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/tests/integration_tests/retrievers/document_compressors/test_base.py b/tests/integration_tests/retrievers/document_compressors/test_base.py new file mode 100644 index 00000000..389d4d04 --- /dev/null +++ b/tests/integration_tests/retrievers/document_compressors/test_base.py @@ -0,0 +1,28 @@ +"""Integration test for compression pipelines.""" +from langchain.document_transformers import EmbeddingsRedundantFilter +from langchain.embeddings import OpenAIEmbeddings +from langchain.retrievers.document_compressors import ( + DocumentCompressorPipeline, + EmbeddingsFilter, +) +from langchain.schema import Document +from langchain.text_splitter import CharacterTextSplitter + + +def test_document_compressor_pipeline() -> None: + embeddings = OpenAIEmbeddings() + splitter = CharacterTextSplitter(chunk_size=20, chunk_overlap=0, separator=". ") + redundant_filter = EmbeddingsRedundantFilter(embeddings=embeddings) + relevant_filter = EmbeddingsFilter(embeddings=embeddings, similarity_threshold=0.8) + pipeline_filter = DocumentCompressorPipeline( + transformers=[splitter, redundant_filter, relevant_filter] + ) + texts = [ + "This sentence is about cows", + "This sentence was about cows", + "foo bar baz", + ] + docs = [Document(page_content=". ".join(texts))] + actual = pipeline_filter.compress_documents(docs, "Tell me about farm animals") + assert len(actual) == 1 + assert actual[0].page_content in texts[:2] diff --git a/tests/integration_tests/retrievers/document_compressors/test_chain_extract.py b/tests/integration_tests/retrievers/document_compressors/test_chain_extract.py new file mode 100644 index 00000000..0fcfebf9 --- /dev/null +++ b/tests/integration_tests/retrievers/document_compressors/test_chain_extract.py @@ -0,0 +1,36 @@ +"""Integration test for LLMChainExtractor.""" +from langchain.chat_models import ChatOpenAI +from langchain.retrievers.document_compressors import LLMChainExtractor +from langchain.schema import Document + + +def test_llm_chain_extractor() -> None: + texts = [ + "The Roman Empire followed the Roman Republic.", + "I love chocolate chip cookies—my mother makes great cookies.", + "The first Roman emperor was Caesar Augustus.", + "Don't you just love Caesar salad?", + "The Roman Empire collapsed in 476 AD after the fall of Rome.", + "Let's go to Olive Garden!", + ] + doc = Document(page_content=" ".join(texts)) + compressor = LLMChainExtractor.from_llm(ChatOpenAI()) + actual = compressor.compress_documents([doc], "Tell me about the Roman Empire")[ + 0 + ].page_content + expected_returned = [0, 2, 4] + expected_not_returned = [1, 3, 5] + assert all([texts[i] in actual for i in expected_returned]) + assert all([texts[i] not in actual for i in expected_not_returned]) + + +def test_llm_chain_extractor_empty() -> None: + texts = [ + "I love chocolate chip cookies—my mother makes great cookies.", + "Don't you just love Caesar salad?", + "Let's go to Olive Garden!", + ] + doc = Document(page_content=" ".join(texts)) + compressor = LLMChainExtractor.from_llm(ChatOpenAI()) + actual = compressor.compress_documents([doc], "Tell me about the Roman Empire") + assert len(actual) == 0 diff --git a/tests/integration_tests/retrievers/document_compressors/test_chain_filter.py b/tests/integration_tests/retrievers/document_compressors/test_chain_filter.py new file mode 100644 index 00000000..1068a1e6 --- /dev/null +++ b/tests/integration_tests/retrievers/document_compressors/test_chain_filter.py @@ -0,0 +1,17 @@ +"""Integration test for llm-based relevant doc filtering.""" +from langchain.chat_models import ChatOpenAI +from langchain.retrievers.document_compressors import LLMChainFilter +from langchain.schema import Document + + +def test_llm_chain_filter() -> None: + texts = [ + "What happened to all of my cookies?", + "I wish there were better Italian restaurants in my neighborhood.", + "My favorite color is green", + ] + docs = [Document(page_content=t) for t in texts] + relevant_filter = LLMChainFilter.from_llm(llm=ChatOpenAI()) + actual = relevant_filter.compress_documents(docs, "Things I said related to food") + assert len(actual) == 2 + assert len(set(texts[:2]).intersection([d.page_content for d in actual])) == 2 diff --git a/tests/integration_tests/retrievers/document_compressors/test_embeddings_filter.py b/tests/integration_tests/retrievers/document_compressors/test_embeddings_filter.py new file mode 100644 index 00000000..15a13e39 --- /dev/null +++ b/tests/integration_tests/retrievers/document_compressors/test_embeddings_filter.py @@ -0,0 +1,39 @@ +"""Integration test for embedding-based relevant doc filtering.""" +import numpy as np + +from langchain.document_transformers import _DocumentWithState +from langchain.embeddings import OpenAIEmbeddings +from langchain.retrievers.document_compressors import EmbeddingsFilter +from langchain.schema import Document + + +def test_embeddings_filter() -> None: + texts = [ + "What happened to all of my cookies?", + "I wish there were better Italian restaurants in my neighborhood.", + "My favorite color is green", + ] + docs = [Document(page_content=t) for t in texts] + embeddings = OpenAIEmbeddings() + relevant_filter = EmbeddingsFilter(embeddings=embeddings, similarity_threshold=0.75) + actual = relevant_filter.compress_documents(docs, "What did I say about food?") + assert len(actual) == 2 + assert len(set(texts[:2]).intersection([d.page_content for d in actual])) == 2 + + +def test_embeddings_filter_with_state() -> None: + texts = [ + "What happened to all of my cookies?", + "I wish there were better Italian restaurants in my neighborhood.", + "My favorite color is green", + ] + query = "What did I say about food?" + embeddings = OpenAIEmbeddings() + embedded_query = embeddings.embed_query(query) + state = {"embedded_doc": np.zeros(len(embedded_query))} + docs = [_DocumentWithState(page_content=t, state=state) for t in texts] + docs[-1].state = {"embedded_doc": embedded_query} + relevant_filter = EmbeddingsFilter(embeddings=embeddings, similarity_threshold=0.75) + actual = relevant_filter.compress_documents(docs, query) + assert len(actual) == 1 + assert texts[-1] == actual[0].page_content diff --git a/tests/integration_tests/retrievers/test_contextual_compression.py b/tests/integration_tests/retrievers/test_contextual_compression.py new file mode 100644 index 00000000..60eb206b --- /dev/null +++ b/tests/integration_tests/retrievers/test_contextual_compression.py @@ -0,0 +1,25 @@ +from langchain.embeddings import OpenAIEmbeddings +from langchain.retrievers.contextual_compression import ContextualCompressionRetriever +from langchain.retrievers.document_compressors import EmbeddingsFilter +from langchain.vectorstores import Chroma + + +def test_contextual_compression_retriever_get_relevant_docs() -> None: + """Test get_relevant_docs.""" + texts = [ + "This is a document about the Boston Celtics", + "The Boston Celtics won the game by 20 points", + "I simply love going to the movies", + ] + embeddings = OpenAIEmbeddings() + base_compressor = EmbeddingsFilter(embeddings=embeddings, similarity_threshold=0.75) + base_retriever = Chroma.from_texts(texts, embedding=embeddings).as_retriever( + search_kwargs={"k": len(texts)} + ) + retriever = ContextualCompressionRetriever( + base_compressor=base_compressor, base_retriever=base_retriever + ) + + actual = retriever.get_relevant_documents("Tell me about the Celtics") + assert len(actual) == 2 + assert texts[-1] not in [d.page_content for d in actual] diff --git a/tests/integration_tests/test_document_transformers.py b/tests/integration_tests/test_document_transformers.py new file mode 100644 index 00000000..d5a23dba --- /dev/null +++ b/tests/integration_tests/test_document_transformers.py @@ -0,0 +1,31 @@ +"""Integration test for embedding-based redundant doc filtering.""" +from langchain.document_transformers import ( + EmbeddingsRedundantFilter, + _DocumentWithState, +) +from langchain.embeddings import OpenAIEmbeddings +from langchain.schema import Document + + +def test_embeddings_redundant_filter() -> None: + texts = [ + "What happened to all of my cookies?", + "Where did all of my cookies go?", + "I wish there were better Italian restaurants in my neighborhood.", + ] + docs = [Document(page_content=t) for t in texts] + embeddings = OpenAIEmbeddings() + redundant_filter = EmbeddingsRedundantFilter(embeddings=embeddings) + actual = redundant_filter.transform_documents(docs) + assert len(actual) == 2 + assert set(texts[:2]).intersection([d.page_content for d in actual]) + + +def test_embeddings_redundant_filter_with_state() -> None: + texts = ["What happened to all of my cookies?", "foo bar baz"] + state = {"embedded_doc": [0.5] * 10} + docs = [_DocumentWithState(page_content=t, state=state) for t in texts] + embeddings = OpenAIEmbeddings() + redundant_filter = EmbeddingsRedundantFilter(embeddings=embeddings) + actual = redundant_filter.transform_documents(docs) + assert len(actual) == 1 diff --git a/tests/unit_tests/test_document_transformers.py b/tests/unit_tests/test_document_transformers.py new file mode 100644 index 00000000..26354aef --- /dev/null +++ b/tests/unit_tests/test_document_transformers.py @@ -0,0 +1,15 @@ +"""Unit tests for document transformers.""" +from langchain.document_transformers import _filter_similar_embeddings +from langchain.math_utils import cosine_similarity + + +def test__filter_similar_embeddings() -> None: + threshold = 0.79 + embedded_docs = [[1.0, 2.0], [1.0, 2.0], [2.0, 1.0], [2.0, 0.5], [0.0, 0.0]] + expected = [1, 3, 4] + actual = _filter_similar_embeddings(embedded_docs, cosine_similarity, threshold) + assert expected == actual + + +def test__filter_similar_embeddings_empty() -> None: + assert len(_filter_similar_embeddings([], cosine_similarity, 0.0)) == 0 diff --git a/tests/unit_tests/test_math_utils.py b/tests/unit_tests/test_math_utils.py new file mode 100644 index 00000000..34b390a5 --- /dev/null +++ b/tests/unit_tests/test_math_utils.py @@ -0,0 +1,39 @@ +"""Test math utility functions.""" +from typing import List + +import numpy as np + +from langchain.math_utils import cosine_similarity + + +def test_cosine_similarity_zero() -> None: + X = np.zeros((3, 3)) + Y = np.random.random((3, 3)) + expected = np.zeros((3, 3)) + actual = cosine_similarity(X, Y) + assert np.allclose(expected, actual) + + +def test_cosine_similarity_identity() -> None: + X = np.random.random((4, 4)) + expected = np.ones(4) + actual = np.diag(cosine_similarity(X, X)) + assert np.allclose(expected, actual) + + +def test_cosine_similarity_empty() -> None: + empty_list: List[List[float]] = [] + assert len(cosine_similarity(empty_list, empty_list)) == 0 + assert len(cosine_similarity(empty_list, np.random.random((3, 3)))) == 0 + + +def test_cosine_similarity() -> None: + X = [[1.0, 2.0, 3.0], [0.0, 1.0, 0.0], [1.0, 2.0, 0.0]] + Y = [[0.5, 1.0, 1.5], [1.0, 0.0, 0.0], [2.0, 5.0, 2.0]] + expected = [ + [1.0, 0.26726124, 0.83743579], + [0.53452248, 0.0, 0.87038828], + [0.5976143, 0.4472136, 0.93419873], + ] + actual = cosine_similarity(X, Y) + assert np.allclose(expected, actual)