Two-Stage RAG Chatbot: OpenAI & Supabase Vector Search

This n8n workflow builds a sophisticated two-stage document retrieval system integrated with OpenAI and Supabase vector store. It starts by querying file metadata descriptions via a Supabase RPC function to identify the most relevant files based on user input similarity scores. Only these top-matching files are then used for deeper vector similarity searches on document chunks, drastically reducing noise and improving retrieval precision.

The benefits include optimized performance for handling hundreds of files split into chunks, minimizing token limits in AI prompts, and delivering highly relevant answers. It solves common RAG pitfalls like irrelevant data flooding LLMs, leading to better accuracy, faster queries, and cost savings on API calls.

Ideal for developers, AI enthusiasts, and data engineers managing knowledge bases, FAQs, or content libraries in apps. Use cases span customer support chatbots, internal document search, publishing platforms, and agentic AI systems needing scalable, precise retrieval without full re-indexing.