Legal-Knowledge RAG Assistant
A Greek-language assistant that answers questions over a large body of legislation — hybrid retrieval (keyword + vector), LLM reranking, and a multi-layer verifier so answers stay grounded in the actual text, with citations.
The challenge
Legislation is dense, cross-referenced and unforgiving — a wrong or invented answer is worse than no answer. Staff and the public need fast, accurate answers grounded strictly in the real legal text, in Greek.
What we built
We built a retrieval-augmented assistant that combines keyword (BM25) and semantic (vector) search with reciprocal-rank fusion, LLM-based query understanding and reranking, and a multi-layer anti-hallucination verifier that checks every answer against the retrieved sources before it is shown. Separate user, engineer and admin portals plus a public, searchable codex of the law.
- Hybrid retrieval (BM25 + vector) with rank fusion
- LLM query understanding + reranking
- Multi-layer anti-hallucination verifier (answers cite the source text)
- Greek-language, domain-tuned
- User / engineer / admin portals + public law codex