Medical RAG Chatbot — Diabetic Retinopathy

🔗 Live Demo: Try it on HuggingFace Spaces

A retrieval-augmented generation (RAG) chatbot that answers natural-language questions about diabetic retinopathy, grounded in open-access research papers. It retrieves relevant passages from a document collection and uses a large language model to generate accurate, source-cited answers.

This project extends my Diabetic Retinopathy Grading work (a vision model that grades DR severity) into the LLM / NLP domain, adding RAG, vector databases, embeddings, and LLM API integration.

⚠️ Research demo only — not medical advice. Answers are generated from a limited document set and may be incomplete or wrong. Always consult a qualified ophthalmologist.

---

How RAG works (in brief)

A plain LLM answers from memory and can hallucinate. RAG instead grounds the answer in real documents:

1. Ingest — documents are split into chunks and converted to vectors (embeddings) stored in a vector database.
2. Retrieve — at question time, the question is embedded and the most semantically similar chunks are pulled from the database.
3. Generate — those chunks are given to the LLM, which is instructed to answer only from them and cite its sources.

PDFs ──▶ chunk ──▶ embed ──▶ ChromaDB        (ingestion, run once)
                                  │
Question ──▶ embed ──▶ retrieve top-k chunks ─┘
                                  │
                  chunks + question ──▶ LLM ──▶ cited answer   (at query time)

---

Stack

Component	Choice
Orchestration	LangChain
Vector store	ChromaDB (persistent, on disk)
Embeddings	sentence-transformers/all-MiniLM-L6-v2 (free, runs on CPU)
LLM	Llama 3.3 70B via Groq API (free tier)
UI	Gradio
Deployment	HuggingFace Spaces

---

Project structure

Medical-RAG-Chatbot/
├── app.py                  # Gradio chat UI (HF Spaces entry point)
├── requirements.txt
├── .env.example            # template for your GROQ_API_KEY
├── .gitignore
├── data/                   # source PDFs (gitignored)
└── src/
    ├── config.py           # all settings in one place
    ├── download_sources.py # fetch open-access DR papers from arXiv
    ├── ingest.py           # load → chunk → embed → store in ChromaDB
    └── rag_pipeline.py     # retrieve → prompt Groq → cited answer

---

Setup and run

# 1. Install dependencies
pip install -r requirements.txt

# 2. Add your free Groq API key
cp .env.example .env
#    then edit .env and paste your key from https://console.groq.com/keys

# 3. Download source documents (open-access arXiv papers on DR)
python -m src.download_sources
#    (or drop your own PDFs into the data/ folder)

# 4. Build the vector store (run once, or after changing documents)
python -m src.ingest

# 5. Launch the chatbot
python app.py
#    open the printed local URL in your browser

---

Example questions

• "What is diabetic retinopathy?"
• "How is the severity of diabetic retinopathy graded?"
• "What deep learning methods are used to detect diabetic retinopathy?"
• "Why is early screening important?"

---

Screenshots

The chatbot answering questions with grounded, source-cited responses, each ending with a safety disclaimer:

---

Configuration

All tunable settings live in src/config.py:

• CHUNK_SIZE / CHUNK_OVERLAP — how documents are split
• EMBED_MODEL — the embedding model
• TOP_K — how many chunks to retrieve per question
• LLM_MODEL / LLM_TEMPERATURE — the Groq model and its creativity

---

Future work

• Add more diseases and clinical guideline sources
• Try a stronger or domain-specific embedding model (e.g. BioBERT-based)
• Add a re-ranking step after retrieval
• Eval