Self-Query: how to provide all possible filter values for AttributeInfo to self-query constructor?

[ronan-zelta]

Checked other resources

• I added a very descriptive title to this question.
• I searched the LangChain documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.

Commit to Help

• I commit to help with one of those options 👆

Example Code

metadata_field_info = [
            AttributeInfo(
                name="competitor_name",
                description=f"Name of competitor company which is associated with document. YOU MUST ONLY choose competitor names from the following "
                            f"list: {competitor_string}. If any other competitor is named, DO NOT include them in the filter, simply put the "
                            f"competitor name in the query field.",
                type="string",
            ),
            AttributeInfo(
                name="feedback_type",
                description=f"Type of customer feedback in the document. Must be one of the following: {feedback_type_string}",
                type="string",
            )
        ]

        doc_contents = ("Short explanations of key insights extracted from customer feedback data (sources include Gong & Zoom call transcripts, "
                        "Intercom & Zendesk tickets, G2 & App Store reviews, etc.)")

        chain = load_query_constructor_runnable(
            ChatOpenAI(model="gpt-4-turbo", temperature=0, openai_api_key=self.config['OpenAI']['API_KEY]),
            doc_contents,
            metadata_field_info
        )

        examples = [ "...", "...", ...]

        prompt = get_query_constructor_prompt(
            doc_contents, metadata_field_info, examples=examples
        )

        result = chain.invoke({"query": raw_query})

Description

My RAG application allows users to apply filters & type a query to find relevant documents (in my case these documents are related to customer feedback). I am trying to use LangChain self-querying to extract filter values from the prompt (e.g. if the user types "positive feedback about X", without applying any filters, I want to be able to extract filters from the prompt and apply them automatically when querying the DB)

Because I'm using the LangChain output to filter my DB, I need the extracted filters to be exact string matches, and don't want to extract filters which do not exist.

I was hoping there would be some sort of possible values parameter where I could provide a list of possible filter values which the self-query engine could choose from. But I haven't been able to find anything like this.

At the moment I'm just listing the possible options in the AttributeInfo description, and I plan to have some logic after the filter generation to remove any filters which are not exact matches. But it'd be a lot better if I could just provide the string options to the self-query engine in the first place to avoid this workaround.

• Am I correct in saying that there's no way to do this at the moment?
• Is my current workaround the best approach available?

Would appreciate any help or advice!

System Info

langchain==0.1.16
langchain-community==0.0.33
langchain-core==0.1.43
langchain-openai==0.1.3
langchain-text-splitters==0.0.1

[dosubot[bot]]

You're right in your understanding that LangChain's self-query engine currently does not support a direct way to specify a list of possible filter values for AttributeInfo within the query construction process. This means that the approach you're considering, where you list possible options in the AttributeInfo description and then filter out non-matches post-generation, remains the best available method.

The AttributeInfo is used to inform the query constructor about the metadata fields that can be queried, but it doesn't include functionality for restricting those fields to a predefined list of values. This limitation requires implementing additional logic outside of the LangChain query construction to ensure the extracted filters match your exact requirements.

For more detailed insights into how AttributeInfo and the query constructor work, you might want to explore the SelfQueryRetriever class in the LangChain repository, particularly the from_llm class method. This method demonstrates how metadata_field_info is processed and might offer some inspiration for implementing your filtering logic.

If you have any more questions or need further assistance, feel free to ask.

Sources

• libs/langchain/langchain/chains/query_constructor/prompt.py
• libs/langchain/langchain/retrievers/self_query/base.py

About Dosu

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.

[Anna-Bianca]

I am having the same problem, that you, I was wondering how did you managed to solve it?