07

AI assistant for your documents

AI assistant for your documents: answers that cite their sources.

An assistant that answers from your internal documents, cites the passages it used and says when the information is missing. A controlled alternative to using consumer tools on sensitive data.

This work fits your situation if…

  • Scattered information

    Procedures, contracts, meeting minutes, technical documentation: the answer exists, but you have to know where to look. The same questions keep coming back to the same people.

  • A consumer tool used on internal documents

    Staff paste documents into an online assistant to save time. You want an equivalent tool whose sources, access and hosting you control.

  • Answers that must be verifiable

    In your line of work, an answer without a source cannot be used. The assistant has to show where each element comes from and say when it does not know.

What the engagement can deliver.

The scope and deliverables are agreed together during scoping.

A defined, prepared corpus
Document selection, update rules, access rights and chunking suited to search. Users know what the assistant covers and what it does not.
Cited answers and explicit refusals
Every answer points to the passages used. Outside its corpus, the assistant says so instead of inventing, as the Archiviste, on the French version of the Work page, does for this site’s case studies.
An evaluation on your own questions
A set of real questions, with the expected answers and sources, is used to measure quality before go-live and after each change to the corpus.
An architecture suited to your documents
A model hosted in Europe, an approved cloud service or a model deployed in your environment: the choice follows the sensitivity of the documents and your operating constraints.

A clear sequence.

  1. Choose a useful scope

    Start from one audience, one family of documents and frequent questions. A narrow, well-maintained scope is better than an exhaustive corpus that is poorly maintained.

  2. Build and evaluate

    Prepare the corpus, build search and citations, then measure the answers against the question set, including the questions the assistant must decline to answer.

  3. Open it up, then widen

    Make the assistant available to a first group, track the questions left unanswered and decide which documents to add.

Case studies that show the work.

Plaidoria — sourced legal drafting

Every reference in a legal brief is checked against official sources, Légifrance (the official French legal database) and Judilibre (the French courts’ case-law database); a reference that cannot be found blocks generation. The same principle guides a document assistant: no answer without a verifiable source.

Read the case study : Plaidoria — sourced legal drafting

Before we start.

Is this a chatbot?

It is a conversational assistant, but limited to your documents and required to cite its sources. It does not answer on topics outside its corpus and does not replace a person when a decision is at stake.

Do our documents leave our environment?

That depends on the architecture chosen. With a model deployed in your own environment, the documents stay on your infrastructure; with an external service, the data sent, how long it is retained and the provider’s commitments are examined before choosing.

Can we see an example?

The Archiviste, on the French version of this site’s Work page, is a public example: it answers questions about the case studies and cites every passage it uses. For your case, a limited first scope lets you evaluate the answers on your own questions.

Which documents can be used?

Mainly text documents: procedures, contracts, reports, documentation, meeting minutes. Scanned documents require a character recognition (OCR) step; tables and diagrams are examined case by case.

06

Next step

Let’s start from your need.

Describe the documents involved, the people who consult them and three questions they often ask. That starting point is enough to assess feasibility.

Discuss your project