In production

An internal assistant that answers with the source in view

Natural-language questions about management-system procedures, the technical archive and financial documentation, each answer citing the document it came from.

Let’s talk about your case
Sector
Power generation
Timeline
23 weeks
Team
3 people from Recursa Labs
Status
In production, with monthly support

The context

A power generation company runs on living, scattered documentation: quality, safety and environmental procedures, drawings and equipment manuals, financial records. It all exists — finding the exact paragraph is what takes time, and it usually depends on whoever has been there longest.

The client authorized us to show their logo, not the detail of the project. So we tell the case without naming them and without internal figures.

The real problem

The typical question isn't hard to answer, it's hard to locate: which procedure sets the deadline to report an incident, which drawing matches a piece of equipment, what a six-month-old invoice actually says.

An assistant that answers beautifully but makes things up is worse than no assistant. From day one the requirement was traceability: every answer had to be checkable against its source document, and access had to respect who can see what.

How we approached it

  1. An assistant over the management system

    We started with quality, safety and environmental procedures. The assistant answers in natural language and shows the document and page it drew from, so the person can verify it.

  2. Reading the technical archive

    Drawings and manuals go through optical recognition tuned for technical documents — rotated text, symbols, embedded tables. Semantic indexing was adapted to the content instead of a generic chunking strategy.

  3. Queries over financial documentation

    For financial data, natural-language questions translated into SQL behind a security layer: sanitization, an allow-list of operations and injection prevention, plus validation of XML invoices.

  4. Quality measured, not promised

    We built a suite of precision and edge-case tests with the client, with evaluation rubrics their team reviews. Every prompt or configuration change is evaluated before it reaches users.

  5. Operations with explicit commitments

    Ongoing service includes first response to critical incidents in under an hour, infrastructure cost monitoring, and recalibration when the model drifts.

What is running today

  • An assistant in production, available from Microsoft Teams and the internal web.
  • Answers that cite the source document, subject to each repository's permissions.
  • Three fronts solved in 23 weeks: management system, technical archive and financial documentation.
  • An evaluation set owned by the client, which survives any change of model.
  • Monthly support with first response to critical incidents in under an hour.

What it is built with

Azure AI FoundryAzure AI SearchAzure AI Document IntelligenceMicrosoft TeamsSharePointCosmos DBText-to-SQLPython
From idea to implementation

Let's talk about your operation.

Tell us the process, the systems involved and the outcome you need. We'll assess the scope and the next steps with you.

Request an assessmentSee how we work

Services involved

AI agents and assistants

Document assistants and AI agents wired into your operation, with permissions and quality criteria.

Data and analytics

Put your data sources in order and turn them into dashboards, indicators and queries you can trust.

Request an assessment