abdelmoniim/tunisian-labor-law-ai-assistant

AI assistant answering questions about Tunisian labor law (Code du Travail), built with hybrid retrieval (BM25 + embeddings), reranking, and Groq/Llama # ⚖️ Tunisian labor law ai assistant An AI assistant that answers questions about the Tunisian Labor Law (Code du Travail, 2016 edition) in French, grounded strictly in the official text — no external knowledge, no hallucinated articles. Built from scratch with a custom hybrid-retrieval RAG pipeline — deliberately without LangChain, in order to understand what happens under the hood at every stage (chunking, embeddings, hybrid retrieval, reranking, generation) — as a portfolio project demonstrating retrieval-augmented generation fundamentals end to end. > ⚠️ This is a technical demonstration project and does **not** constitute legal advice. ## How it works ``` PDF (Code du Travail 2016) │ ▼ Ingestion → parses raw text into structured articles │ (article number, livre/titre/chapitre/section, │ modification history) ▼ Chunking → builds two indexes from the parsed articles: │ • BM25 (lexical/keyword index) │ • ChromaDB (dense vector index, bge-m3 embeddings) ▼ Retrieve → hybrid search at query time: │ 1. BM25 search + vector search (top 20 each) │ 2. Reciprocal Rank Fusion (RRF) merges both lists │ 3. Cross-encoder reranks the fused candidates │ 4. Top 5 most relevant articles are kept ▼ Generate → the retrieved articles are passed as context to │ Groq/Llama, which answers strictly from that │ context and cites the article numbers used ▼ Gradio chat UI ``` ### Why hybrid retrieval? - **BM25** catches exact legal terminology and article-number matches that embeddings can miss. - **Dense vectors (bge-m3)** catch semantically related questions phrased differently from the text. - **RRF** combines both ranked lists without needing to tune a weighting scheme. - **Cross-encoder reranking (bge-reranker-base)** does a final, more expensive but more accurate pass over the fused candidates before they reach the LLM. ### Grounding & anti-hallucination The system prompt instructs the model to answer *only* …