ORIA Lab · Document assistant

An answer, its source — or nothing.

Eight internal documents from a fictional microfinance institution. Ask your question: the relevant passage is found, highlighted and cited with its reference. If no source matches, the demo says so.

100% local, nothing is sent ≈ 0.1 MB downloaded on first run

How it works

  1. Chunking

    Each document is split into sections, which keep their heading and the document reference.

  2. Normalisation

    Your question is lower-cased, stripped of accents and stop words, reduced to word stems and enriched with business synonyms.

  3. Ranking

    The BM25 algorithm ranks sections by how often, and how distinctively, the search terms appear.

  4. Confidence threshold

    Below a matching threshold, no answer is shown: “I don’t know” beats an invented answer.

What it means for your business

In production, the assistant combines semantic search with a language model to write an answer, always with its sources, over your real documents and with role-based access rights. Quality is measured on a reference question set defined with your teams.

Enterprise document assistant
Search runs in your browser: your questions are neither sent nor stored. The documents and the organisation are fictional, created for this demo.

Frequently asked questions

Is this generative AI?
Not in this demo: it is a lexical search (BM25) that returns an exact excerpt of the document. It illustrates the principle behind our assistants — a verifiable answer with its source. In production, a language model writes the answer from these passages.
Why do some questions get no answer?
Either the information is not in the documents, or the question uses words too far from those of the procedures. Semantic search reduces the second case; the first must remain a non-answer.
Can we use it with our own documents?
Yes: that is what the Document assistant service covers — your procedures, contracts and archives indexed securely, with a measured 4 to 6-week pilot.

Other demos

100% local · free

Document → data

A scanned invoice becomes structured, cross-checked data in seconds, without leaving your browser.

Launch the demo

100% local · free

Weak-signal detection

Spot, among 300 fictional customers, those slowly disengaging — earlier than a classic threshold rule.

Launch the demo

Next step

45 minutes to identify your first win.

A structured conversation with an AI architect — no commitment, no jargon.

Book my free assessment