title	Medical RAG Chatbot
emoji	⚕️
colorFrom	blue
colorTo	green
sdk	docker
app_port	8501

Medical Research RAG Chatbot

An enterprise-grade Retrieval-Augmented Generation (RAG) chatbot designed to ingest, process, and query complex medical research papers (e.g., PubMed, ArXiv). It uses local embedding models to ensure data privacy and Google Gemini 1.5 Flash for high-quality, hallucination-free generation.

🚀 Project Overview

Medical Research RAG Chatbot | Python, LangChain, FAISS, Docker Live: huggingface.co/spaces/arinShr/medical-rag-assistant

• Built a Retrieval-Augmented Generation pipeline with LangChain, Gemini 1.5 Flash, and a local FAISS vector store, including multi-hop query decomposition for comparative queries across documents.

🛠️ Tech Stack

• Document Ingestion: PyMuPDF, LangChain (RecursiveCharacterTextSplitter - 512 tokens / 50 overlap)
• Embeddings & Vector Store: sentence-transformers/all-MiniLM-L6-v2, FAISS (Local)
• LLM & Chain: Google Gemini 1.5 Flash, LangChain RetrievalQA
• Frontend: Streamlit (with confidence scores & source citations)
• Evaluation: RAGAS (Faithfulness, Answer Relevancy, Context Precision)
• Deployment: Docker, GitHub Actions

🧠 Advanced Architecture: Multi-Hop Query Decomposition

Standard single-hop RAG often fails on comparative queries across multiple documents because the embedded query vector heavily skews toward the vocabulary of a single document. To solve this, the pipeline implements Query Decomposition:

1. Uses a heuristic (RegEx) to detect comparative queries (compare, differences, etc.)
2. Utilizes the LLM to split the query into independent sub-queries.
3. Retrieves chunks independently for each sub-query to guarantee both source documents are represented.
4. Synthesizes a comprehensive answer from the deduplicated, merged context.

Robust Rate Limit Handling

Because query decomposition initiates multiple parallel LLM calls, it can easily exhaust API quotas (429 RESOURCE_EXHAUSTED). The system uses tenacity-style exponential backoff decorators to silently catch limit exceptions, sleep, and auto-retry, ensuring stable performance under load.

⚠️ Technical Learnings

• Chunk Boundary Sensitivity: Dense clinical prose with separated entities (e.g., a drug name and its numeric efficacy separated by ~100 characters of explanatory text) can easily fall victim to "semantic splitting" if chunk overlap is too small. Initial configurations with a 50-character overlap cleanly severed key clinical statistics from their contextual entities, causing false-negative retrievals. Bumping CHUNK_OVERLAP to 150 mathematically guaranteed these long-range numeric relationships remained co-located in the vector space, successfully restoring retrieval accuracy without requiring complex layout parsers.

📂 Project Structure

resume 2/
├── data/
│   ├── raw/                 # Put your medical PDFs here
│   └── vector_store/        # FAISS index is saved here
├── src/
│   ├── config.py            # Global settings
│   ├── ingest.py            # PDF parsing & chunking
│   ├── vector_store.py      # Embeddings & DB building
│   ├── rag_chain.py         # LLM logic
│   └── evaluate.py          # RAGAS metrics
├── app.py                   # Streamlit frontend
├── .env.example             # Template for GEMINI_API_KEY
└── requirements.txt         # Dependencies

⚙️ How to Run Locally

1. Setup Environment

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

2. Configure API Key

Copy the example environment file and add your Google Gemini API key:

cp .env.example .env
# Edit .env and replace with your GEMINI_API_KEY

3. Ingest Data

1. Download 10-20 medical PDFs (e.g., Cardiology, Oncology) and place them in the data/raw/ directory.
2. Run the ingestion and vector store script:

python -m src.vector_store

(This will chunk the PDFs and create a FAISS index in data/vector_store/)

4. Run the Chatbot UI

Launch the Streamlit frontend:

streamlit run app.py

5. Evaluate the Model

To generate the RAGAS evaluation report to validate hallucination prevention (shows engineering rigor):

python -m src.evaluate

🐳 Docker Deployment

To build and run via Docker (e.g., for Hugging Face Spaces):

docker build -t medical-rag .
docker run -p 8501:8501 --env-file .env medical-rag