An AI assistant that answers from your internal documents
Your teams lose time looking for a procedure, a clause or a decision that was already made. A RAG (retrieval-augmented generation) assistant queries your documents, writes an answer and points to the passages it relies on. It still has to respect access rights, and you need a way to measure whether it is right.
The problem
Company knowledge is scattered: an intranet, a shared drive, PDFs, a ticket base, chat tools. Internal search returns documents, not answers, and newcomers depend on the three people who “know”.
A general-purpose language model does not know your documents and makes things up when it does not know. RAG fixes this by handing it, for each question, the relevant excerpts from your base. Answer quality then depends far more on retrieval than on the model.
How we work
-
1
Start from questions, not documents
We collect around fifty real questions from future users, each with the expected answer and the document that holds it. This test set is the reference for the whole project: without it, nobody can say whether a change improves or degrades the assistant.
-
2
Prepare the sources
Text extraction, cleaning, splitting into coherent passages, metadata (date, department, version, confidentiality). Tables, scanned documents and outdated versions are handled explicitly: they cause most wrong answers.
-
3
Build retrieval, then generation
Hybrid search (vector and lexical), result reranking, then an answer that cites its passages. Access rights are enforced at retrieval time: the assistant cannot quote a document the user is not allowed to open.
-
4
Measure, ship, keep it alive
Automatic evaluation on the test set at every change, a log of unanswered questions, one-click user feedback. Syncing with your sources is automated so the assistant does not answer with last year’s procedure.
What makes these projects fail
Indexing everything, unsorted
Three versions of the same procedure in the index give three different answers. Curating sources is a business task, not a technical one, and it needs an owner.
Ignoring access rights
An assistant that summarises an HR document to someone who should not read it is a security incident. Filtering is designed into the architecture, not added after the demo.
Judging on a demo
Ten well-chosen questions always look good. Only an evaluation on real questions, including those whose answer is not in the base, tells you what the system is worth.
What you receive
- Reference question set and evaluation report
- Ingestion and sync pipeline for your sources
- Assistant integrated into your tools (web, intranet, internal chat or API)
- Access-rights filtering and audit log
- Dashboard: questions asked, unanswered questions, user feedback
- Operations documentation and handover to your teams
Frequently asked questions
Do our documents leave the company?
That depends on the architecture, and it is a decision we make with you during scoping. Depending on data sensitivity, we can use a model hosted in Europe under a contractual no-reuse commitment, or an open model deployed on your own infrastructure.
Do we need to train a model on our data?
In the vast majority of cases, no. RAG hands documents to the model at question time, with no retraining. It is simpler to update, easier to audit, and removing a document takes effect immediately.
What happens when the assistant does not know?
It has to say so. We tune and test this behaviour explicitly: when no relevant passage is found, the assistant says so and suggests who to ask, rather than producing a plausible, wrong answer.
Is your situation close to this one?
Describe it in a few lines. We will tell you whether AI is the right answer — and we will also tell you when it is not.
Talk about your project