The context
The firm's consultants spent their days navigating consolidated texts, guidelines and internal notes spread across four overlapping regulations. Keyword search fails on questions phrased in business language, and purely semantic search misses exact references, an article number, a specific recital. The team wanted a single, reliable entry point, hosted in Europe.
The need
An engine able to answer both "Article 28 GDPR" and "what does a processor risk when reusing client data". That requires hybrid retrieval: dense for meaning, lexical for exact references, fused then re-ranked. Every answer had to cite the source passage with its official reference, a non-negotiable condition for a firm whose credibility rests on accuracy. And the entire system had to remain within a European perimeter.
What we delivered
A search engine built on Weaviate, which natively combines vector search and BM25. A versioned corpus of 30,000 pages: consolidated texts, guidelines, selected case law. Multilingual embeddings computed locally, generation by Mistral hosted on OVHcloud, re-ranking before answering. A FastAPI API exposes the search to the firm's internal tools, and every answer returns the exact paragraph, its official reference and a confidence level. This foundation prepares the next step: plugging analysis agents, such as our AI Act agent, into the same base.