LangGraph Agent & pgvector RAG Optimization

Hey! We're a small SaaS team building an AI assistant that helps us answer customer security questionnaires using our internal documents. The app is already built with Next.js, FastAPI, OpenAI, and Supabase. It works pretty well for basic questions, but we're stuck on two annoying issues. First, the assistant sometimes pulls answers from old policy documents instead of the latest versions. For example, when we ask about our data retention policy, it occasionally uses a document from last year and cites the wrong page. Second, we have a LangGraph workflow that saves generated answers to our database. When an API request times out, the agent sometimes repeats the action and creates duplicate records. We've tried a few fixes, but haven't been able to solve these properly. Here's what we'd like help with: Make our pgvector retrieval prioritize current document versions. Fix source citations so they point to the correct document sections. Debug the LangGraph workflow and prevent duplicate database writes. Add some basic tracing so we can see why a tool call fails. Test the changes against our existing sample questions. We don't need new features or a full rebuild. Just someone who understands RAG and agent workflows well enough to identify the issues and fix them without breaking what already works. Our stack is Python, FastAPI, LangGraph, OpenAI, PostgreSQL/pgvector, Supabase, and Next.js. We have the code on GitHub and a staging environment ready. This should be a fairly small job for someone who's dealt with similar AI problems before. If this sounds familiar, please tell us briefly how you'd approach it.