Context and architecture
Built as an ENSIA NLP course project, the system addresses the cross-lingual and relational structure of Algerian legal research.
The retrieval layer combines dense vector similarity, lexical BM25, and graph traversal so an answer can follow references between articles, decrees, and related legal entities.
A Groq-hosted Llama 3.3 70B generation layer uses the retrieved context to produce RTL-aware Arabic and French responses.
Core capabilities
- Bilingual French and Arabic retrieval
- Hybrid vector, BM25, and graph search
- Entity-relationship context for generation
- RTL-aware React interface
Engineering challenge
The key challenge was preserving legal cross-references while handling Arabic and French spelling variation. The solution combines language-aware preprocessing with multiple retrieval signals instead of relying on a single embedding score.
Outcome
Ranked first in the ENSIA academic cohort.