Medical Research Assistant for Doctors

Project Description

An AI-powered medical assistant designed for doctors to ask clinical questions and receive grounded, evidence-based answers using Retrieval-Augmented Generation (RAG) and hybrid LLM fallback orchestration.

Medical PDFs placed in the data/ directory are automatically extracted, chunked, embedded, indexed in ChromaDB, and retrieved via semantic similarity search to minimize hallucinations. Answers are generated using Llama 3.1-8B via Groq API with OpenAI GPT-4o-mini fallback for robustness.

---

🧠 How it Works

Ingestion:
Medical documents in data/ are auto-loaded, text-extracted, chunked (1500 characters with 300 overlap), embedded using a Sentence Transformer model, and stored in a persistent ChromaDB vector index.

Retrieval:
Doctor queries are converted into embeddings and matched against the vector index using semantic similarity search. The system retrieves Top 5 most relevant medical literature chunks with metadata and relevance scores.

Hybrid Generation:

• Primary: Generates answers using Llama 3.1-8B via Groq API, strictly grounded in retrieved medical context
• Fallback: Automatically switches to OpenAI (GPT-4o-mini) if Groq is unavailable or fails
• If retrieved documents lack sufficient information, the model responds with a clear medical disclaimer instead of hallucinating

---

🧩 Tech Stack

• LangChain: RAG orchestration framework
• ChromaDB: Vector storage and semantic retrieval
• Hugging Face Embeddings: sentence-transformers/all-MiniLM-L6-v2
• LLM APIs: Groq (Llama 3.1-8B-Instant) + OpenAI (GPT-4o-mini fallback)
• PDF Processing: PyMuPDF (fitz)

---

✨ Key Features

Clinical Context Priority: Answers only from retrieved medical PDFs
Hybrid RAG: Falls back to OpenAI if Groq fails
Source Attribution: Retrieved medical sources displayed in UI + logs
Optimized Retrieval: 1500-char chunks for full clinical meaning
Drug Safety Alerts: Highlights drug interactions, adverse effects, contraindications
Robust LLM Fallback: Auto-switches to OpenAI if Groq fails
Utilities: python-dotenv, tqdm, numpy, pandas (optional for eval scripts).

---

🧠 Core Modules Included

Module	Purpose
ingest_data.py	Extracts text from PDFs → splits → embeds → stores in ChromaDB
verify_index.py	Validates vector index integrity and tests similarity search
rag_pipeline.py	Embeds doctor query → retrieves top 5 chunks → calls LLM APIs with fallback
evaluate_rag.py	Benchmarks retrieval quality + answer grounding using RAG metrics
app.py	Streamlit UI for doctor interaction + source preview + safety alerts

---

Images

Image1

Image2

Image3

Image4

Image5

Installation and Setup

1️⃣ Clone the Repository

    git clone https://github.com/Shivp45/MedIntel.git  
    cd MedIntel

2️⃣ Activate your virtual environment

    python -m venv ai_service  
    ai_service\Scripts\activate

3️⃣ Install dependencies

    pip install -r requirements.txt

4️⃣ Add medical PDFs
Copy your medical research papers, guidelines, toxicology, drug docs into:

    data/*/*

5️⃣ Run document ingestion (creates vector DB + index)
Only if you add new data file/files...

    python backend/ingest_data.py   

Expected output should show:

    ✔ 32 PDFs detected  
    ✔ ~4688 chunks embedded  
    ✔ Vector DB stored in embeddings/chroma/ (with chroma.sqlite3 inside it)

6️⃣ (Optional) Verify vector index is working

    python verify_index.py

This will test semantic search like:

    "What are common medical treatments?" → returns 3 similar chunks

7️⃣ Run the doctor chat UI

    streamlit run app.py

The app will open automatically in your browser at:

    http://localhost:8501

Flow Diagram

                ┌──────────────────────┐
                │  data/ (PDF files)   │
                └──────────┬───────────┘
                           ▼
                ┌──────────────────────┐
                │  ingest_data.py      │
                │  - Extract text      │
                │  - Chunk docs        │
                │  - Generate embeds   │
                │  - Store in Chroma   │
                └──────────┬───────────┘
                           ▼
                ┌────────────────────────────┐
                │ embeddings/chroma/ (DB)   │
                │  Vector Index Stored Here  │
                └──────────┬────────────────┘
                           ▼
                ┌──────────────────────┐
                │  rag_pipeline.py     │
                │  - Embed query       │
                │  - Retrieve top K    │
                │  - Send context → LLM│
                └──────────┬───────────┘
                           ▼
     ┌────────────── Groq API (Primary LLM) ───────────────┐
     │                Model: Llama 3.1-8B                  │
     └───────────────┬───────────────────────┬─────────────┘
                     │  (if fails 401/timeout)│
                     ▼                       │
     ┌──────────── OpenAI API (Fallback LLM) │
     │           Model: GPT-4o-mini          │
     └────────────┬──────────────────────────┘
                  ▼
                ┌──────────────────────┐
                │  app.py (Streamlit UI)│
                │  - Show answer        │
                │  - Show sources       │
                │  - Safety alerts      │
                └──────────┬───────────┘
                           ▼
                ┌──────────────────────┐
                │ evaluate_rag.py      │
                │ - Retrieval metrics  │
                │ - Keyword grounding  │
                └──────────┬───────────┘
                           ▼
                ┌──────────────────────┐
                │ verify_index.py      │
                │ - Test vector search │
                │ - Check DB integrity │
                └──────────────────────┘

[Doctor Query] → [Embed Query] ↓ [Retrieve Top 5 Chunks from ChromaDB] ↓ [Send Context + Query to Groq (Llama 3.1-8B)] ↓ {If fails} ↓ [OpenAI Fallback (GPT-4o-mini)] ↓ [Return Answer + Sources + Safety Warnings] ↓ [Display in Streamlit UI]