Description
We tried building our AI support workflow in n8n but hit a wall. We need someone who can build it properly in Python instead. What we need: a backend service that handles customer queries, decides whether to route them to Claude or OpenAI depending on the type of question, pulls relevant answers from our internal documentation (RAG-style, we have ~200 docs), and returns a response with a confidence score. If confidence is low, it should trigger an escalation instead of guessing. This needs to be real, maintainable code — not a no-code workflow. We specifically want: FastAPI backend with a clean API layer Custom routing logic between Claude API and OpenAI API based on query type (not just a static rule — something that can be tuned over time) RAG pipeline using a vector database (open to Pinecone, Weaviate, or your recommendation) Confidence scoring tied to retrieval quality, not just LLM self-reported certainty Proper error handling and retry logic — this needs to run reliably, not just work in a demo Basic logging/analytics so we can see what's routed where and why We tried the no-code route first and it fell apart once we needed branching logic and reliable error handling, so please don't pitch us an n8n/Make solution — we specifically need custom Python. What we're looking for: Strong Python backend experience (FastAPI or similar) Real production experience with Claude API and/or OpenAI API, not just tutorials RAG/vector search experience Comfortable explaining architectur