Medical RAG

A production-ready, extensible Retrieval-Augmented Generation (RAG) system with advanced features for multi-lingual, and domain-specific (especially medical) question answering.

---

🚀 Key Features

• Hybrid Retrieval: Combines vector search (semantic) and keyword/BM25 search for robust, accurate retrieval.
• Multi-modal Support: Handles text, images (via GPT-4o), audio (Whisper), and video (transcript + frame chunking).
• Advanced Chunking: Intelligent, semantic, and structural chunking for optimal context retrieval.
• Query Expansion & Self-Query: LLM-based query rewriting, metadata extraction, and multi-perspective retrieval.
• Reranking: LLM-based, cross-encoder, and MMR reranking for high-quality context selection.
• Conversational Memory: Short-term (session) and long-term (user profile, semantic search) memory for multi-turn dialogue.
• Medical Knowledge Integration: UMLS, NER, and translation pipeline for Vietnamese/English medical Q&A.
• Evaluation & Monitoring: RAGAS, LLM-based evaluation, and experiment tracking (CometLLM, WandB).
• Finetuning: Supports embedding and LLM finetuning (LoRA, SentenceTransformers).
• Scalable Serving: Supports OpenAI API, LMDeploy, vLLM, Infinity, and self-hosted Qdrant.
• Extensible UI: Streamlit-based, with session management, chunking preview, and medical mode.

---

🏗️ System Architecture

User Query (VN/EN/Multimodal)
    ↓
[Translation] (if needed)
    ↓
[Medical NER] → [UMLS API] → [MMR Ranking] → [Medical Context]
    ↓
[Query Expansion/Self-Query]
    ↓
[Hybrid Retrieval: Qdrant (vector+BM25)]
    ↓
[Reranking (LLM/Cross-Encoder/MMR)]
    ↓
[Context Merging: Memory + Medical + Retrieved]
    ↓
[LLM Generation]
    ↓
[Evaluation & Monitoring]
    ↓
[UI/Feedback]

---

⚡ Quick Start

1. Prerequisites

• Docker & docker-compose
• (Optional) UMLS API key for medical features

2. Build & Run

docker-compose up --build

• Access the UI at: http://localhost:8501

3. Qdrant Setup

Qdrant is used as the vector database. It is automatically started via docker-compose.
For manual setup or advanced options, see: https://qdrant.tech/documentation/quick-start/

---

🧠 Advanced RAG Techniques

• Query Expansion: LLM generates multiple queries for broader semantic coverage.
• Self-Query: Extracts metadata (tags, filters) from user queries for precise retrieval.
• Hybrid Search: Combines vector and keyword search for best coverage.
• Reranking: Uses LLMs or cross-encoders to select the most relevant context.
• Multi-stage Retrieval: Supports multi-hop, agent-based, and memory-augmented retrieval.

---

🏥 Medical Q&A Pipeline

• Vietnamese/English Support: Automatic translation (EnViT5) for Vietnamese queries.
• Medical NER: Extracts diseases, drugs, symptoms, etc.
• UMLS Integration: Retrieves medical knowledge and relations, reranked by MMR.
• Context Merging: Medical knowledge is prioritized and merged with document retrieval for LLM generation.

---

🧩 Extensibility

• Chunking: Plug-and-play strategies (semantic, structural, fast, intelligent).
• Memory: Modular short-term and long-term memory, easy to extend for new use cases.
• Retrieval: Swap or combine retrieval backends (Qdrant, Elastic, etc.).
• LLM Serving: Compatible with OpenAI API, LMDeploy, vLLM, Infinity, etc.

---

📊 Evaluation & Monitoring

• RAGAS: Automated metrics for context relevancy, recall, answer similarity, and correctness.
• LLM-based Evaluation: Use GPT-4o or other LLMs for qualitative assessment.
• Experiment Tracking: CometLLM, WandB integration for prompt and model monitoring.

---

🔒 Security & Best Practices

• API Keys & Secrets: Never commit .env, settings.py, or any API keys (see .gitignore).
• Finetuning Data: All /finetuning folders are ignored by git.
• User Data: Session/user data is not stored unless explicitly enabled.

---

📚 References

• Qdrant Documentation
• SentenceTransformers
• RAGAS
• UMLS
• LMDeploy
• vLLM
• Infinity

---

📝 Contribution & Development

• Fork, clone, and develop in feature branches.
• PRs are welcome! Please do not commit any API keys, .env, or /finetuning data.

---

💡 Example Use Cases

• Medical Q&A (Vietnamese/English)
• Enterprise document search
• Multi-modal knowledge assistant
• Research assistant with memory and context

---

For more details, see the documentation in each module and the medical/README.md for medical pipeline specifics.