All case studies / Case study 01 / 12
Norvelle Group
“Atlas”, RAG document assistant
Atlas: an assistant that knows all your documents
In short
Procedures, contracts, HR notes, emails: Atlas indexes every internal document by meaning and answers teams’ questions in plain words, citing its sources every time.
- Client
- Norvelle Group
- Sector
- Business services
- Size
- 200 staff, 3 offices
- A few secondsinstead of long searches through shared folders
- 100%of answers cite their sources
- 0documents leaving your infrastructure
The context
At Norvelle, the information exists but sleeps in thousands of files: procedures in a shared folder, HR rules as PDFs, contracts in the document system, decisions lost in emails. Newcomers keep asking the same questions, and managers keep answering them by hand.
The challenge
- 01
Keyword search finds nothing when you don’t know the right term.
- 02
Documents exist in several versions, and no one knows which one is current.
- 03
An answer without a source is worthless to HR or legal.
- 04
The documents are internal and must stay on the company’s infrastructure.
Our approach
Every document is split into passages, and each passage becomes a vector that captures its meaning, stored in PostgreSQL with pgvector. For each question, Atlas finds the passages closest in meaning, respecting the user’s access rights, then a language model writes a short answer citing the original document and page. The model can run on your servers, or through an API hosted in Europe.
The solution, step by step
- 01
Your documents
PDFs, Word files, emails, wiki: we gather them without taking them out of your infrastructure.
- 02
Indexed by meaning
Each passage becomes a point in a space where related ideas cluster together.
- 03
A question, in plain words
Your teams ask their question as they would ask a colleague.
- 04
The right passages
Atlas finds the most relevant excerpts, not just keywords.
- 05
An answer with sources
The answer cites its sources: you can always check.
Results
- −60%repetitive questions to managers and HR
- A few secondsto find the right procedure
- 100%of answers with source and page
- 1 weeksaved onboarding a newcomer
What we deliver
- Connectors to the shared drive, document system and email
- Semantic chunking and indexing pipeline
- PostgreSQL + pgvector with access rights
- Q&A assistant with citations
- Automatic updates when a document changes
Technologies
- Python
- LangChain
- PostgreSQL + pgvector
- Open-source LLM or API
- Docker
Timeline
6 to 10 weeks
Team
1 AI engineer, 1 developer
Illustrative cases: clients and figures are fictional, representative of the solutions we deliver.