Control the number of resources documents to return (zero or more).

[onalv]

I'm using pinecone as vector database for a Q&A app (for documents). I want to control how many resource documents are returned but I want it to be variable, so that if there're no relevant vectors, just return 0 source documents, but if there are more return 2 or 3 or 4, maybe using like a filter for the probability of how related is it to the question or just a config value? (e.g. probability of text being related to the question >= 0.5). Currently I can pass a number to asRetriever but this doesn't comply with what I need.

const chain = ConversationalRetrievalQAChain.fromLLM(
    llm,
    vectorStore.asRetriever(3),
    {
      memory,
      returnSourceDocuments: true,
      qaChainOptions: {
        type: "stuff",
        prompt,
      },
      questionGeneratorChainOptions: {
        llm: nonStreamingModel,
        template: CONDENSE_TEMPLATE,
      },
    }
  );

Do I need to use a different config or different way of doing it?

[dosubot[bot]]

🤖

Hello,

Based on the context provided, it seems you want to variably control the number of returned resource documents based on their relevance to a given query. This is indeed possible by modifying the getRelevantDocuments function in the SelfQueryRetriever class.

Unfortunately, the exact code for the SelfQueryRetriever class is not provided in the context. However, I can guide you on a general approach to achieve this.

The getRelevantDocuments function currently does not take any parameter to limit the number of returned documents. You can add a parameter, say limit, to this function and use it to control the number of documents returned.

If the getRelevantDocuments function uses the Pinecone vector database to retrieve the documents, you can use the limit parameter of the Pinecone query function to limit the number of returned documents. The Pinecone query function usually looks like this:

index.query( vectors, { limit: 10 } )

In this case, you can replace 10 with the limit parameter of the getRelevantDocuments function.

Please note that this is a general approach and the exact implementation may vary based on the actual code of the SelfQueryRetriever class and the Pinecone library.

For more information, you can refer to the examples/src/retrievers/pinecone_self_query.ts file in the LangChainJS repository.

Additionally, there was a similar feature request in the past: Feature request: Customizable document retrieval parameters (k and score threshold) for Index-related chains with VectorStore integration. The feature was implemented in a pull request by a user named mayt. You can refer to the pull request here for more details.

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

Sources

• Feature request: Customizable document retrieval parameters (k and score threshold) for Index-related chains with VectorStore integration
• examples/src/retrievers/pinecone_self_query.ts

---

This response is meant to be useful and save you time. 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.

If you want to continue the conversation, start your reply with @dosu-bot.

[bracesproul]

Hey @onalv I haven't tested this so I'm not 100% sure, but you should try passing an instance of ScoreThresholdRetriever as your retriever. The ScoreThresholdRetriever class allows for returning n results that are within a certain predefined score threshold.