How We Built Arabic AI Law: RAG + BM25 + Whisper in Production
Hybrid retrieval for Arabic legal text
Pure vector search missed exact article references. We combined Pinecone similarity with BM25 keyword ranking and merge scores before passing context to GPT-4o.
Voice and RTL
Whisper handles Arabic audio queries; the React frontend is fully RTL with accessible typography for long legal passages.
Monetization
Stripe subscription tiers gate advanced models and case workspace limits. Admin dashboard toggles prompts and model selection without redeploys.
Case workspaces
Users organize research per legal matter, isolating embeddings and chat history from other cases — critical for firm confidentiality.