sanaaloute/La-loi-dit

A specialized AI platform that cover west africa countries laws and regulations from justice to commerce, from traffic to family code, in short all regulations, policies and law. # Burkina Faso Legal AI A production-grade **agentic AI legal research assistant** for Burkina Faso law and regulations (national law, OHADA law, case law, official gazette). It answers legal questions in French (and other languages) with verified, traceable citations to real legal sources. Design priorities, in order: **accuracy, reliability, explainability, traceability, low hallucination**. The system refuses to guess: every substantive statement must trace to a retrieved evidence chunk, fabricated citations are rejected automatically, and low-confidence answers are flagged or escalated for human review. > **Legal disclaimer.** This system is a legal *research* aid. Its answers are > grounded in the cited sources but do **not** constitute legal advice. Always > consult a licensed legal professional for your specific situation. ## Features | Capability | How it works | |---|---| | Multi-agent reasoning pipeline | 13-node LangGraph workflow: guardrail → planner → context → memory → retrieval → conflict resolution → evidence ranking → reasoning → reflection → citation verification → response → output guardrail | | Grounded answers | Responses are composed strictly from retrieved evidence; every claim carries a numeric citation verified against the evidence list | | Hybrid retrieval | Vector (Milvus) + keyword (BM25) + government/regulation/case-law/news workers, fused with RRF and authority-weighted ranking | | Legal conflict resolution | Official source > newer amendment > timeline-in-force; unresolvable conflicts are surfaced, never silently resolved | | Authority-weighted ranking | Constitution > OHADA treaty > amended law > law > decree > … > blog (see `backend/core/constants.py`) | | Anti-hallucination controls | Citation verification, reflection self-critique, bounded retries (max 1), grounded-answer policy, refusal path | | Memory | MemGPT-style tiers: short-term buffer, summaries, long-term semantic memory, user preferences | | Durable conversations | T …