How to perform a fuzzy search across conversations when using LangGraph’s AsyncPostgresSaver as a checkpointer?

Product Help LangGraph
1

Hey everyone,

I’ve been using LangGraph (served through a FastaAPI) for a while to serve my assistant to multiple users, and I think I’m using its abstractions in the right way (but open to roasts). In my implementation to persist chat history I use AsyncPostgresSaver as a checkpointer for my Agent:

graph = workflow.compile(checkpointer=AsyncPostgresSaver(self._pool))

My thread_id is a string composed of the user ID plus the date. That way, when I want to list all conversations for a certain user, I run something like:

SELECT
    thread_id,
    metadata -> 'writes' -> 'Generate Title' ->> 'title' AS conversation_title,
    checkpoint_id
FROM checkpoints
WHERE metadata -> 'writes' -> 'Generate Title' ->> 'title' IS NOT NULL
  AND thread_id LIKE '%%{user_id}%%';

Now i got the thread_id and can display all the messages like this

config: Dict[str, Any] = {"configurable": {"thread_id": thread_id}}
state = await agent.aget_state(config)
messages = state[0]["messages"]
return messages

Note: for me a thread is basically a chat with a title, what you would normally see on the left bar of ChatGPT.

The problem:

Now I want to search inside a conversation.

The issue is that I’m not 100% sure how the messages are actually stored in Postgres. I’d like to run a string search (or fuzzy search) across all messages of a given user, then group the results by conversation and only show conversations that match.

My questions are:

  • Can this be done directly using the AsyncPostgresSaver storage format, or would I need to store the messages in a separate, more search-friendly table?

  • Has anyone implemented something like this with LangGraph?

  • What’s the best approach to avoid loading every conversation into memory just to search?

  • Cause i can see stuff is saved as Binary Data sometimes (which makes sense for documents)? But I cannot believe that the text part of a message is not searchable

Any advice or patterns you’ve found useful would be appreciated!

2

Hi! We don’t support accessing the checkpoint data in Postgres outside of the utilities the LangGraph library provides you. For your use case, it is best to store the conversation data in a separate schema you’ve designed.