🏥 MediQuery

RAG-Powered Medical Document Q&A using LangChain, ChromaDB & Groq

Built with the tools and technologies:

---

🧠 Project Summary

MediQuery is an AI-powered medical document intelligence assistant that answers drug-related questions strictly from official FDA documentation — with source citations and confidence scores. Built using Retrieval-Augmented Generation (RAG), it combines local HuggingFace embeddings with Groq's blazing-fast LLM inference for zero-hallucination, citation-backed answers.

🔗 Try it Live on Streamlit

---

🚀 Features

• 📄 FDA-Grounded Answers
  Responses are sourced exclusively from official FDA drug label PDFs — no hallucination, no guesswork.
• 🔎 Source Citations & Confidence Scores
  Every answer includes the exact source document, page number, and a High / Medium / Low relevance rating per retrieved chunk.
• 💬 Conversational Memory
  Two-step rephrase-then-retrieve pipeline resolves follow-up questions using full chat history context.
• 📤 PDF Upload at Runtime
  Upload your own medical PDFs and query them instantly — indexed on the fly.
• 🛡️ No Hallucination
  Strict prompt engineering with temperature=0. If the answer isn't in the documents, the model says so.
• 📊 RAG Evaluation Pipeline
  Evaluated using an LLM-as-judge methodology (equivalent to RAGAS) across 20 medical queries.

---

🏗️ Architecture

User Question
      ↓
Rephrase to standalone question (if follow-up)
      ↓
ChromaDB Vector Search (top-4 chunks by cosine similarity)
      ↓
HuggingFace Embeddings (all-MiniLM-L6-v2 — runs locally)
      ↓
Context + Chat History → Groq LLM (Llama 3.1 8B Instant)
      ↓
Answer + Source Citations + Confidence Scores

---

💊 Knowledge Base

8 official FDA drug labels pre-loaded — 272 pages, 984 chunks:

Drug	Category
Metformin	Diabetes
Amoxicillin	Antibiotic
Atorvastatin	Cholesterol
Ibuprofen	Pain Relief
Lisinopril	Blood Pressure
Sertraline	Antidepressant
Warfarin	Blood Thinner
Omeprazole	Acid Reflux

---

📊 Evaluation Results

Evaluated using a custom LLM-as-judge pipeline across 20 queries covering all 8 drug labels. Llama 3.1 8B via Groq API served as the evaluation judge.

Metric	Score
Faithfulness	0.82 / 1.00
Answer Relevancy	0.78 / 1.00

• Faithfulness — answers grounded in retrieved context with no hallucination.
• Answer Relevancy — answers directly address the question asked.
• 20/20 questions answered. One Sertraline query scored low on relevancy (0.0) — identified as a retrieval gap and future improvement area.

---

🗃️ Project Structure

mediquery-rag/
├── app/
│   ├── main.py                # Streamlit frontend & UI
│   ├── rag_pipeline.py        # RAG logic — vectorstore, LLM chain
│   ├── utils.py
│   └── document_loader.py     # PDF ingestion and chunking
├── data/
│   └── pdfs/                  # FDA drug label PDFs (not committed)
├── tests/
│   └── evaluate_rag.py        # LLM-as-judge evaluation script
├── .streamlit/
│   └── config.toml            # Streamlit configuration
├── .env.example               # API key template
├── requirements.txt           # App dependencies
└── requirements-dev.txt       # Dev + evaluation dependencies

---

🔧 Setup & Installation

Make sure Python 3.12 is installed.

# Clone the repo
git clone https://github.com/Muhammad-Ahmed-Rayyan/mediquery-rag.git
cd mediquery-rag

# Create a virtual environment
python -m venv venv

# On Windows
venv\Scripts\activate
# On macOS/Linux
source venv/bin/activate

# Install app dependencies
pip install -r requirements.txt

# (Optional) Install evaluation dependencies
pip install -r requirements-dev.txt

---

🔑 API Configuration

• ⚙️ Groq API Key — .env

cp .env.example .env
# Edit .env and paste your GROQ_API_KEY

Get a free API key at console.groq.com.

• 📄 FDA Drug Label PDFs — data/pdfs/

Download drug label PDFs from DailyMed and place them in data/pdfs/.

---

▶️ Running the App

cd app
streamlit run main.py

The vectorstore is built automatically on first run (~2 minutes).

---

🧪 Running Evaluation

python tests/evaluate_rag.py

Runs 20 test questions across all 8 drug labels and outputs faithfulness and answer relevancy scores. Results are saved to tests/ragas_results.csv.

---

🔮 Future Improvements

• Semantic chunking instead of fixed 800-token chunking
• Cross-encoder re-ranking for better retrieval precision
• Conversation history summarization to manage context window growth
• Docker containerization for portable deployment
• Expanded knowledge base with more drug categories

---

💬 Demo Questions to Try

• "What are the contraindications of Metformin for kidney patients?"
• "Can Warfarin and Ibuprofen be taken together?"
• "What is the recommended dose of Amoxicillin for children?"
• "What are the serious risks of taking Warfarin?"
• "What are the side effects of Atorvastatin?"

---

⭐ Don’t forget to star the project on GitHub!