Retrieval & RAG
RAG Playground
Watch retrieval happen: dense + keyword + reranking, then a grounded answer with citations.
Production proof: The retrieval core of my enterprise RAG chatbot and 92%-accuracy legal Q&A. The engine runs on your device; nothing is uploaded. Read the case study →
⚡ Run it live with your own API key no key set
Your key is held in memory for this session (kept for this tab only if you tick the box) and sent only to the provider you choose, directly from your browser. This site has no backend server — you can confirm it in your browser's Network tab: requests go only to the provider, never to this site. Live runs use small, fast models with hard token caps; a typical run costs well under $0.01, billed to your own account.
Honest limits: no website can protect a key from a compromised browser or malicious extension. Use a key with a spending cap, and revoke it after playing if you like.
How it’s engineered
- Retrieval quality, not the language model, is what makes RAG trustworthy. This demo shows the whole retrieval stack running on your device.
- Dense retrieval (MiniLM embeddings) captures meaning; BM25 lexical retrieval catches exact terms like error codes and IDs. Hybrid fusion (reciprocal rank fusion) combines both, which consistently beats either alone.
- A reranker re-scores the fused shortlist so the best passage lands first; the answer then cites the exact chunks it used, so every claim is auditable.
- Toggle the retrieval methods to watch the ranking change: the "HTTP 429" query rewards lexical search, "get my money back" rewards dense search. That gap is exactly why production RAG is hybrid.