Live demo, chat with your files

Ask your case files a question. See the exact page it came from.

Every answer cites the document and the page it came from, with the clause marked on it. If the answer is not in the files, it says so. That guard runs on the server, not on the model's goodwill.

The suggested questions are saved runs of the real pipeline, so they answer instantly and use no tokens. Type your own and it goes to the live model.

The live model hit today's free-tier limit. The suggested questions still work.
How it works

Retrieval first, and a hard stop before the model

The question is matched against the pages by meaning, not keywords. If nothing scores high enough, the server answers "not in the documents" on its own. The model never gets a chance to invent an answer.

Your question
→
Rank pages by meaning
→
Hard cutoff, on the server
→
Answer, with the page
In your business

Where this fits

Whole folders, not three files

Point it at a Drive folder or a case room with thousands of pages. Scanned documents get OCR first, so the answers stay page-exact.

A real vector database

This demo keeps the math in memory. At scale it swaps in a proper vector store, so search stays fast across huge document sets.

Your formats, your hosting

Word and Outlook exports too, not just PDF. For sensitive files it can run in your own cloud or the region you choose (US, EU, wherever your data must stay), with nothing leaving it.

Built with
Gemini embeddings Semantic search (cosine, in-memory) Page-exact citations Server-enforced "not in the docs" guard Nothing stored

Got a stack of contracts or reports to interrogate?

Tell me what you would want to ask them. I will wire this to your real documents, with the citation guard on.

Get in touch →