youssefrekikk/Insurance-Conversational-Agent-Tunisian-Dialect

🇹🇳 Tunisian Insurance Chatbot A dialect-aware, multimodal RAG chatbot for Tunisian insurance queries Features • Architecture • Getting Started • Model Training • Docs --- ## 📸 Screenshots 📱 Gallery: Agentic Workflow & Observability | Complex Query | User Profile | |:---:|:---:| | | | | Document Upload | Observability Trace (Jaeger) | |:---:|:---:| | | | --- ## 🎯 Overview This project delivers a production-ready conversational AI assistant tailored for the **Tunisian insurance market**. Unlike generic chatbots, it natively understands: - **Tunisian Darija** — The spoken dialect mixing Arabic, French, and Berber - **Arabizi** — Latin-script encoding of Arabic (`3` = ع, `7` = ح, `9` = ق) - **Code-switching** — Seamless transitions between languages mid-sentence > *Example:* `"Aandi sinistre fel pare-choc, chnowa l'procédure?"` → The system understands this as an insurance claim inquiry about bumper damage. --- ## 🙏 Acknowledgments - This project was developed as part of a **Projet Tutoré** at SUP'COM (Higher School of Communication of Tunis) under the supervision of: **Firas Ben Hassan** — Program Owner of AllianzGPT at Allianz Technology, Munich. His expertise in enterprise-scale Generative AI systems provided invaluable guidance for this project's architecture and implementation. - Unsloth — Efficient fine-tuning framework - Qdrant — High-performance vector database - OpenRouter — Unified LLM API - The Tunisian NLP community for dialect resources --- ## ✨ Features ### 🗣️ Dialect-Native Processing - Custom **Qwen3-8B adapter** trained on Tunisian corpus - Real-time Arabizi normalization (`karhba` → `vehicle`) - Translation layer mapping dialect terms to formal insurance vocabulary ### 🔍 Intelligent RAG Pipeline - **Hybrid Retrieval**: Dense vectors (BGE-M3) + Sparse search (BM25) + Episodic memory - **HyDE Strategy**: Uses fine-tuned LLM to generate hypothetical answers for better retrieval - **Cross-Encoder Rerank …