Interactive demo · RAG

A production RAG assistant, on your documents.

Pick a question. Watch the assistant retrieve relevant chunks, generate an answer, and cite its sources. The flow is real. The data is from a sample knowledge base.

Interactive

From question to cited answer.

This is a scripted walkthrough that runs entirely in your browser. The architecture, retrieval steps, and citation pattern are exactly how we build them in production. Want to see it on your data?

Knowledge Assistant

Pick a question to see the assistant retrieve, reason, and cite.

Try a question
How it works

Four stages, end to end.

01

Source documents are ingested

Confluence, Notion, PDFs, help center, and wiki content flow through a nightly job. Pages are chunked along semantic boundaries with source URLs preserved.

02

Hybrid retrieval finds the right context

A user question hits both vector search (embeddings) and keyword search (BM25). Results are reranked. The top eight chunks pass to the model.

03

Claude answers with citations

Claude 3.5 is prompted to answer only from retrieved context and to refuse low-confidence questions. Every answer cites the source chunks inline.

04

Evaluation catches regressions

A Langfuse-backed eval harness runs nightly on a curated gold set. Any accuracy drop blocks the next deploy.

Stack

Tools used in this demo.

Claude 3.5 Pinecone Voyage embeddings BM25 Langfuse Vercel Edge Next.js
// contact REPLYING IN 24 HOURS

Have a problem worth shipping a real system for?

Tell us what you're trying to build. We reply within 24 hours.