Ask questions. Get cited answers. From your own documents.
We deploy retrieval-augmented chat over the documents your team already has — contracts, case files, patient notes, manuals, SOPs — running entirely on hardware you control. Every answer points back to the source, and nothing is sent to a public API.
Start with a $999 pilot — limited to our first 5 customers (then $1,999). No production data leaves your network — or you don't continue.
What we deploy
- Hybrid search (dense + keyword) with reranking for accuracy
- A cited source for every claim — no unverifiable answers
- Role-based access controls and query logging
- Local model runtime — works with the internet cut
- Answers in English or Vietnamese, whatever language the source is written in
See it in action
What it looks like in practice
The same architecture, applied across very different teams — your industry is just the data source.
A law firm asks across case files and discovery: "What did the witness say about the delivery date?" — and gets the cited passage.
A clinic searches visit notes and charts: "When was this patient last prescribed amoxicillin?" — PHI never leaves the building.
A CPA firm asks across IRS forms and publications: "Can a sole proprietor deduct a home office on Schedule C?" — the same pipeline you can try on this site right now.
A manufacturer queries vendor contracts and equipment manuals: "Which suppliers have a 30-day termination clause?"
A records room digitizes twenty years of scanned files — OCR first, then every page becomes searchable and citable.
Any team turns its internal wiki and SOPs into an assistant new hires can actually ask questions.
Try it on this site: 2,885 IRS forms
Our demo runs the production version of this pipeline — hybrid search over every IRS form, instruction and publication plus 100 synthetic W-2 records, with a citation on every claim. Leave a name and email, then ask it anything about federal tax forms and check each source yourself.
See it live: Private Document Assistant
A working proof of concept: chat with a library of 1,500+ research papers — hybrid search, a cited source for every claim, and answers generated entirely on local hardware. The same architecture works over your documents.
What the 30-day pilot looks like
One workflow, your real documents, and a measured result before you commit to anything larger.
Week 1: we audit a sample of your documents — file types, scans, access rules — and agree on 25 real questions with known answers.
Weeks 2–3: ingestion, hybrid search and the local model go live on your hardware for a small group of users.
Week 4: we score the system against those 25 questions, show you every miss, and hand over the numbers — you decide whether to expand.
Common questions
How is this different from just uploading files to ChatGPT?
Three ways. The documents stay on hardware you control instead of being sent to a third party. The system searches your whole corpus rather than the handful of files in one conversation. And every answer cites the source passage, so a reviewer can verify it instead of trusting it.
How many documents can it handle?
Tens of thousands on a single machine. Our public demo runs over 1,500 research papers, and the architecture scales well past that — the limit in practice is the quality of your documents, not the count.
What file types work?
PDFs, Word documents, spreadsheets, emails, plain text, and web pages. Scanned PDFs need an OCR pass first; we check for that before quoting, because it is the most common hidden cost in a document project.
Can different staff see different documents?
Yes. Access control is per user or per role, applied at retrieval time — someone without permission on a file gets no answers drawn from it, not a redacted version of one.
What happens when the answer isn't in the documents?
It says so. A system that guesses when it has nothing to cite is worse than useless in a professional setting, so we tune for refusal over invention and test that behaviour explicitly during the pilot.
Can it answer in Vietnamese when the documents are in English?
Yes. The question, the documents and the answer do not have to share a language — a Vietnamese question over English contracts gets a Vietnamese answer that cites the English passage, so a reviewer can check the original wording.
Related use cases
Proven in regulated industries
We've packaged this capability for compliance-heavy fields — see how it lands in yours.
Want this running on your hardware?
Book a 30-minute discovery call. We'll scope a $999 pilot — limited to our first 5 customers (then $1,999) — no data leaves your network, or you don't continue.