n8n community support agent
A deployed support agent for the n8n community: Claude answers from the official docs with citations, and anything it isn't sure about escalates to a human review queue.
// try it live
This is the real deployed tool, not a recording. Answers are AI-generated from the official n8n docs and may contain errors.
// the problem
n8n's community answers the same setup and configuration questions over and over, while the docs' keyword search can't handle questions asked in plain English — and an AI answering unsupervised would confidently make things up.
// the approach
- –Ingested n8n's full public documentation (1,311 pages, 9,632 chunks) into a Supabase pgvector store, with the source URL preserved on every chunk.
- –An n8n workflow orchestrates each question: webhook intake → semantic search → Claude drafts an answer with a structured confidence field → the workflow branches on it.
- –The safety gate is code, not prompt engineering: an answer only auto-replies if it cites a source that was actually retrieved — Claude's own claimed sources are verified against the search results, never trusted.
- –Low-confidence answers never reach the user; they land in a review queue with Claude's draft and reasoning attached. A daily request budget caps API spend before any paid call is made.
// the outcome
Live and answering real questions: high-confidence answers arrive with clickable citations into the official docs, unanswerable questions get refused and escalated instead of guessed at, and every exchange is logged with confidence, sources, and cost.