How to use a persistent vector store as the docstore in a MultiVectorRetriever setup

[latifs]

Trying to pass a persistent vector Store as a docstore in a MultiVectorRetriever setup:
Since I wanted to persist the inMemoryVectorStore just in case.
how could I either transform my full_text_vector store to a inMemory one

import os
from langchain.retrievers.multi_vector import MultiVectorRetriever
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from pinecone import Pinecone
from langchain_pinecone import PineconeVectorStore
from pydantic import BaseModel
from langchain_core.runnables import RunnableLambda, RunnableParallel

embeddings = OpenAIEmbeddings(model="text-embedding-3-large")

# User input
class ChatInput(BaseModel):
    question: str
    namespace: str

# LLM
llm = ChatOpenAI(temperature=0, model="gpt-4o-2024-08-06")

# Pinecone initialization
pc = Pinecone(api_key=os.environ.get("PINECONE_API_KEY"))
index = pc.Index(os.environ.get("PINECONE_INDEX"))

def create_multivector_retriever(namespace):
    summaries_vectorstore = PineconeVectorStore(index=index, embedding=embeddings, namespace=f"{namespace}_summaries")
    full_text_vectorstore = PineconeVectorStore(index=index, embedding=embeddings, namespace=f"{namespace}_full")
    
    return MultiVectorRetriever(
        vectorstore=summaries_vectorstore,
        docstore=full_text_vectorstore,
        id_key="doc_id",
    )

# Prompt template
template = """Answer the question based only on the following context, which can include text and tables:
{context}
Question: {question}
"""
prompt = ChatPromptTemplate.from_template(template)

# RAG pipeline
def retrieve_and_generate(inputs):
    retriever = create_multivector_retriever(inputs["namespace"])
    context = retriever.get_relevant_documents(inputs["question"])
    
    return {
        "context": context,
        "question": inputs["question"]
    }

_inputs = RunnableParallel(
    {
        "question": lambda x: x["question"],
        "namespace": lambda x: x["namespace"],
    }
).with_types(input_type=ChatInput)

chain = _inputs | RunnableLambda(retrieve_and_generate) | prompt | llm | StrOutputParser()

[dosubot[bot]]

I found some similar discussions that might help:

1. Is there a way to set the vectorstore as the docstore when setting up a langchain multivector retriever? ^[1]
2. Persistent directory storage for docstore in MultiVectorRetriever ^[2]
3. MultiVector Retriever along with document embedding ^[3]

To use a persistent vector store as the docstore in a MultiVectorRetriever setup, you can follow this approach:

1. Initialize the Vector Stores: Use PineconeVectorStore for both the summaries and full text, as you have done.

2. Setup the Retriever: Pass the PineconeVectorStore instances to the MultiVectorRetriever.

Here's a simplified example:

from langchain_pinecone import PineconeVectorStore
from langchain.retrievers.multi_vector import MultiVectorRetriever

# Initialize Pinecone vector stores
summaries_vectorstore = PineconeVectorStore(index=index, embedding=embeddings, namespace=f"{namespace}_summaries")
full_text_vectorstore = PineconeVectorStore(index=index, embedding=embeddings, namespace=f"{namespace}_full")

# Create MultiVectorRetriever
retriever = MultiVectorRetriever(
    vectorstore=summaries_vectorstore,
    docstore=full_text_vectorstore,
    id_key="doc_id",
)

This setup allows you to use PineconeVectorStore as both the vectorstore and docstore, ensuring persistence. If you need to transform a full_text_vector store to an in-memory one, you can initialize an InMemoryStore and use it similarly ^[4].

To continue talking to Dosu, mention @dosu.

[quartermaine]

@dosu Is there another db we can use for MultiVectorRetriever?