Passing filters to ConversationalRetrievalQAChain with Pinecone DB retriever

[mkorostoff]

I have a function which generates answers, using Pinecone DB as a retriever, which I have based on this document. When generating this answer, only specific documents in the DB should be included. Documents should be included if the metadata field "productID" matches the user input.

It seems like VectorStore.asRetriever supports this, but when I run the code, no filter is applied. The retriever simply returns answers generated from all documents in the index. If I deliberately pass an invalid value like productID: {"$eq": "BAD_DATA"} I still get results, even though no documents have their metadata populated with this value. Here is my code:

async function generateAnswer(vectorStore: PineconeStore, chatMessage: any, productID: string) {
  const model = new ChatOpenAI({});

  const chain = ConversationalRetrievalQAChain.fromLLM(
    model,
    vectorStore.asRetriever(undefined, {
      "productID": {
        "$eq": productID
      }
    }),
    {
      memory: new BufferMemory({
        memoryKey: 'chat_history',
      }),
    },
  );

  const generatednswer = await chain.call({
    question: chatMessage,
  });

  return generatednswer.text;
}

It feels like I'm putting the filter in the wrong place, but I can't find any examples of how to do this. I've considered using PineconeTranslator but I can't see how to do this.

[dosubot[bot]]

Answer generated by a 🤖

Answer

Hello,

Based on the context provided, it seems like you're trying to filter the documents retrieved from the Pinecone DB based on the "productID" metadata field. Unfortunately, the VectorStore.asRetriever method does not directly support filtering based on metadata fields.

However, there's a workaround that you can use. You can extend the VectorStoreRetriever class and override the get_relevant_documents method to apply your filter. Here's an example of how you can do this:

class FilteredRetriever extends VectorStoreRetriever {
  vectorstore: VectorStoreRetriever;
  search_type: string = "similarity";
  search_kwargs: any = {};
  filter_prefix: string;
  
  get_relevant_documents(query: string): Document[] {
    const results = this.vectorstore.get_relevant_documents(query);
    return results.filter(doc => doc.metadata['productID'].startsWith(this.filter_prefix));
  }
}

const filtered_retriever = new FilteredRetriever(vectorstore.asRetriever(), productID);

const chain = ConversationalRetrievalQAChain.fromLLM(
  model,
  filtered_retriever,
  {
    memory: new BufferMemory({
      memoryKey: 'chat_history',
    }),
  },
);

const generatedAnswer = await chain.call({
  question: chatMessage,
});

return generatedAnswer.text;

In this code, FilteredRetriever is a new class that extends VectorStoreRetriever. It overrides the get_relevant_documents method to filter the documents based on the "productID" metadata field.

Please note that this is a workaround and may not be the most efficient solution, especially if you have a large number of documents in your Pinecone DB. The filtering is done after retrieving the documents from the DB, not during the retrieval.

A similar issue was discussed in the LangChain repository: Filtering retrieval with ConversationalRetrievalChain. The solution provided there is similar to the one I've suggested here.

I hope this helps! If you have any further questions, feel free to ask.

Sources

• docs/extras/modules/data_connection/retrievers/how_to/self_query/pinecone.ipynb
• docs/extras/modules/data_connection/retrievers/self_query/pinecone.ipynb
• libs/langchain/langchain/retrievers/self_query/pinecone.py
• libs/langchain/langchain/vectorstores/base.py

---

This response is meant to be useful, save you time, and share context. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.