RAG support assistant
A support chatbot that answers from BarterNow's own policies and docs, falls back across three models when one is slow, and refuses what it can't ground.
- 01Question
- 02Embed
- 03Top 5 sections
- 04Relevance ≥ 0.35?
- 05Model 1 → 2 → 3
- 06Knowledge-base fallback
- 07Answer
The problem
Brands and organisers kept asking the same questions about fees, payouts, deliverables and refunds. Answers lived across policy pages and internal docs, and a generic chatbot would happily make up a refund policy.
What I built
A retrieval-augmented assistant over a curated knowledge base split into company, platform, policy and troubleshooting sections. Questions are embedded, matched against pre-computed section embeddings, and only answered when the best match clears a relevance threshold. Off-topic questions get a short redirect from a lighter model instead of a made-up answer.
Generation tries three models in order with a nine-second budget, and if every model is down it falls back to plain knowledge-base matching, so the widget never goes blank. Embeddings rebuild only when the knowledge changes. The service is internal-only behind the Next.js API, rate-limited per IP, and covered by guardrail tests.
Why it matters for your product
This is the pattern most companies want: an assistant that knows your documents, admits when it doesn't, and keeps working when an AI provider has a bad day. I can build it on your docs, help centre or database.
Want something like this?
Tell me what you're building. I reply within one working day.