Business services

Internal assistant built on the documentation estate

Make a body of procedures and documentation searchable, with sources cited and access rights respected.

Generative AI
RAG
Knowledge management
Access rights

Abstract composition of converging lines, evoking the search for information in a large corpus

Context

An organisation whose operational knowledge — procedures, internal notes, technical documentation — was spread across several storage spaces, with competing versions and differing access perimeters. Teams spent a good deal of time finding the information that applied, and sometimes relied on versions that had expired.

The challenge

To make the corpus searchable in a way that can be trusted: cite the source and its version, respect the access rights already in place, and be able to measure whether answer quality is improving — without creating one more tool that goes stale in its turn.

What AXENEO did

The work we carried out

01
Scoping and corpus
- Qualification of the sources and of the expected uses, audience by audience
- Version management and documentary validity
- Extraction, chunking and enrichment of the metadata
- Carrying over the access rights and partitioning access
02
Designing the assistant
- Retrieval strategy and grounding on the internal sources
- Systematic citation of the document and of its version
- Guardrails, scope of answer and escalation cases
- Integration into the everyday tools
03
Evaluation and industrialisation
- Representative test set and acceptance criteria
- Measurement of relevance and tracking over time
- User feedback loop and updating of the corpus
- Governance framework: permitted uses, data, responsibilities

Results

What the set-up changed

Results are described qualitatively. We do not publish figures we would not be able to document.

Answers with sources, tied to an identified document and to its version.
The documentary access rights already in place are respected.
A test set that makes quality measurable and regressions detectable.
A corpus update cycle built into the documentation practices.
An explicit framework of use, which is what makes extending the perimeter manageable.