Multi-Agent RAG Support System

Overview

Built a multi-agent system in n8n where a central Orchestrator Agent routes user queries to domain-specific Specialist Agents (n8n, Lovable, FlutterFlow). Each specialist queries its own dedicated Supabase Vector Store using OpenAI embeddings, enabling isolated RAG responses per knowledge domain. Ingestion pipeline uses Jina AI to crawl documentation, runs two LLM calls per page to extract relevant links and clean content before indexing, deduplicates content, and loads embeddings into the vector store automatically. PostgreSQL memory persists conversation history per session.

Technical Highlights

• 1Separate vector table per tool — prevents cross-contamination between knowledge bases
• 2Jina AI for clean Markdown extraction from documentation without browser rendering
• 3Two Information Extractor passes per page: links first, then content cleaning
• 4Orchestrator pattern decouples routing from specialist knowledge
• 5Adding a new tool only requires a new ingestion workflow + specialist sub-agent