MultiVector Retriever

It can often be beneficial to store multiple vectors per document. There are multiple use cases where this is beneficial. LangChain has a base MultiVectorRetriever which makes querying this type of setup easy. A lot of the complexity lies in how to create the multiple vectors per document. This notebook covers some of the common ways to create those vectors and use the MultiVectorRetriever.

The methods to create multiple vectors per document include:

Smaller chunks: split a document into smaller chunks, and embed those (this is ParentDocumentRetriever).
Summary: create a summary for each document, embed that along with (or instead of) the document.
Hypothetical questions: create hypothetical questions that each document would be appropriate to answer, embed those along with (or instead of) the document.

Note that this also enables another method of adding embeddings - manually. This is great because you can explicitly add questions or queries that should lead to a document being recovered, giving you more control.

from langchain.retrievers.multi_vector import MultiVectorRetriever

from langchain.document_loaders import TextLoader
from langchain.embeddings import OpenAIEmbeddings
from langchain.storage import InMemoryStore
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.vectorstores import Chroma

loaders = [
    TextLoader("../../paul_graham_essay.txt"),
    TextLoader("../../state_of_the_union.txt"),
]
docs = []
for l in loaders:
    docs.extend(l.load())
text_splitter = RecursiveCharacterTextSplitter(chunk_size=10000)
docs = text_splitter.split_documents(docs)

Smaller chunks

Often times it can be useful to retrieve larger chunks of information, but embed smaller chunks. This allows for embeddings to capture the semantic meaning as closely as possible, but for as much context as possible to be passed downstream. Note that this is what the ParentDocumentRetriever does. Here we show what is going on under the hood.

# The vectorstore to use to index the child chunks
vectorstore = Chroma(
    collection_name="full_documents", embedding_function=OpenAIEmbeddings()
)
# The storage layer for the parent documents
store = InMemoryStore()
id_key = "doc_id"
# The retriever (empty to start)
retriever = MultiVectorRetriever(
    vectorstore=vectorstore,
    base_store=store,
    id_key=id_key,
)
import uuid

doc_ids = [str(uuid.uuid4()) for _ in docs]

# The splitter to use to create smaller chunks
child_text_splitter = RecursiveCharacterTextSplitter(chunk_size=400)

sub_docs = []
for i, doc in enumerate(docs):
    _id = doc_ids[i]
    _sub_docs = child_text_splitter.split_documents([doc])
    for _doc in _sub_docs:
        _doc.metadata[id_key] = _id
    sub_docs.extend(_sub_docs)

retriever.vectorstore.add_documents(sub_docs)
retriever.docstore.mset(list(zip(doc_ids, docs)))

# Vectorstore alone retrieves the small chunks
retriever.vectorstore.similarity_search("justice breyer")[0]

Document(page_content='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\nOne of the most serious constitutional responsibilities a President has is nominating someone to serve on the United States Supreme Court.', metadata={'doc_id': '59899493-92a0-41cb-b6ba-a854730ad74a', 'source': '../../state_of_the_union.txt'})

# Retriever returns larger chunks
len(retriever.get_relevant_documents("justice breyer")[0].page_content)

The default search type the retriever performs on the vector database is a similarity search. LangChain Vector Stores also support searching via Max Marginal Relevance so if you want this instead you can just set the search_type property as follows:

from langchain.retrievers.multi_vector import SearchType

retriever.search_type = SearchType.mmr

len(retriever.get_relevant_documents("justice breyer")[0].page_content)

Summary

Oftentimes a summary may be able to distill more accurately what a chunk is about, leading to better retrieval. Here we show how to create summaries, and then embed those.

import uuid

from langchain.chat_models import ChatOpenAI
from langchain.prompts import ChatPromptTemplate
from langchain.schema.document import Document
from langchain.schema.output_parser import StrOutputParser

chain = (
    {"doc": lambda x: x.page_content}
    | ChatPromptTemplate.from_template("Summarize the following document:\n\n{doc}")
    | ChatOpenAI(max_retries=0)
    | StrOutputParser()
)

summaries = chain.batch(docs, {"max_concurrency": 5})

# The vectorstore to use to index the child chunks
vectorstore = Chroma(collection_name="summaries", embedding_function=OpenAIEmbeddings())
# The storage layer for the parent documents
store = InMemoryStore()
id_key = "doc_id"
# The retriever (empty to start)
retriever = MultiVectorRetriever(
    vectorstore=vectorstore,
    base_store=store,
    id_key=id_key,
)
doc_ids = [str(uuid.uuid4()) for _ in docs]

summary_docs = [
    Document(page_content=s, metadata={id_key: doc_ids[i]})
    for i, s in enumerate(summaries)
]

retriever.vectorstore.add_documents(summary_docs)
retriever.docstore.mset(list(zip(doc_ids, docs)))

# # We can also add the original chunks to the vectorstore if we so want
# for i, doc in enumerate(docs):
#     doc.metadata[id_key] = doc_ids[i]
# retriever.vectorstore.add_documents(docs)

sub_docs = vectorstore.similarity_search("justice breyer")

sub_docs[0]

Document(page_content="The document is a speech given by the President of the United States. The President discusses various important issues and goals for the country, including nominating a Supreme Court Justice, securing the border and fixing the immigration system, protecting women's rights, supporting veterans, addressing the opioid epidemic, improving mental health care, and ending cancer. The President emphasizes the unity and strength of the American people and expresses optimism for the future of the nation.", metadata={'doc_id': '8fdf4009-628c-400d-949c-1d3f4daf1e66'})

retrieved_docs = retriever.get_relevant_documents("justice breyer")

len(retrieved_docs[0].page_content)

Hypothetical Queries

An LLM can also be used to generate a list of hypothetical questions that could be asked of a particular document. These questions can then be embedded

functions = [
    {
        "name": "hypothetical_questions",
        "description": "Generate hypothetical questions",
        "parameters": {
            "type": "object",
            "properties": {
                "questions": {
                    "type": "array",
                    "items": {"type": "string"},
                },
            },
            "required": ["questions"],
        },
    }
]

from langchain.output_parsers.openai_functions import JsonKeyOutputFunctionsParser

chain = (
    {"doc": lambda x: x.page_content}
    # Only asking for 3 hypothetical questions, but this could be adjusted
    | ChatPromptTemplate.from_template(
        "Generate a list of exactly 3 hypothetical questions that the below document could be used to answer:\n\n{doc}"
    )
    | ChatOpenAI(max_retries=0, model="gpt-4").bind(
        functions=functions, function_call={"name": "hypothetical_questions"}
    )
    | JsonKeyOutputFunctionsParser(key_name="questions")
)

chain.invoke(docs[0])

["What were the author's initial areas of interest before college?",
 "What was the author's experience with programming in his early years?",
 'Why did the author switch his focus from AI to Lisp?']

hypothetical_questions = chain.batch(docs, {"max_concurrency": 5})

# The vectorstore to use to index the child chunks
vectorstore = Chroma(
    collection_name="hypo-questions", embedding_function=OpenAIEmbeddings()
)
# The storage layer for the parent documents
store = InMemoryStore()
id_key = "doc_id"
# The retriever (empty to start)
retriever = MultiVectorRetriever(
    vectorstore=vectorstore,
    base_store=store,
    id_key=id_key,
)
doc_ids = [str(uuid.uuid4()) for _ in docs]

question_docs = []
for i, question_list in enumerate(hypothetical_questions):
    question_docs.extend(
        [Document(page_content=s, metadata={id_key: doc_ids[i]}) for s in question_list]
    )

retriever.vectorstore.add_documents(question_docs)
retriever.docstore.mset(list(zip(doc_ids, docs)))

sub_docs = vectorstore.similarity_search("justice breyer")

sub_docs

[Document(page_content='What made Robert Morris advise the author to leave Y Combinator?', metadata={'doc_id': '740e484e-d67c-45f7-989d-9928aaf51c28'}),
 Document(page_content="How did the author's mother's illness affect his decision to leave Y Combinator?", metadata={'doc_id': '740e484e-d67c-45f7-989d-9928aaf51c28'}),
 Document(page_content='What led the author to start publishing essays online?', metadata={'doc_id': '675ccee3-ce0b-4d5d-892c-b8942370babd'}),
 Document(page_content='What measures are being taken to secure the border and fix the immigration system?', metadata={'doc_id': '2d51f010-969e-48a9-9e82-6b12bc7ab3d4'})]

retrieved_docs = retriever.get_relevant_documents("justice breyer")

len(retrieved_docs[0].page_content)

Smaller chunks​

Summary​

Hypothetical Queries​

Smaller chunks

Summary

Hypothetical Queries