Medical QA API

This repository provides a FastAPI-based local service for querying medical documents (PDFs or images) using Google’s Gemini LLM via the LangChain GenAI integration. It ingests a user’s medical document, builds a FAISS vector store, and answers questions with concise bullet-point summaries.

---

📂 Project Structure

medical-qa-api/
├── app/
│   ├── main.py          # FastAPI endpoints and core logic
│   ├── ocr.py           # PDF/image to text extraction
│   └── pipeline.py      # Text splitting, embedding, and FAISS index management
├── data/                # Runtime session data (auto-created)
├── .env                 # Environment variables (e.g. GOOGLE_API_KEY)
├── requirements.txt     # Python dependencies
└── README.md            # This documentation

---

📦 Features

• OCR & Text Extraction

  • PDF → images via pdf2image + text via pytesseract
  • Direct image OCR via pytesseract

• Vector Store

  • Chunks text with CharacterTextSplitter (1000-token chunks, 200 overlap)
  • Embeds with HuggingFaceEmbeddings (all‑MiniLM‑L6‑v2)
  • Indexes embeddings in FAISS, persisted per session

• Retrieval‑Augmented QA

  • Uses Google Gemini (ChatGoogleGenerativeAI) for LLM
  • Custom prompt produces concise bullet‑point summaries

• Session Isolation

  • Each user session stored under ./data/{session_id}
  • Cleanup endpoint removes session data

---

🛠 Requirements

1. Python 3.10+

2. Virtual Environment (recommended)

   python -m venv venv
   source venv/bin/activate

3. Install dependencies

   pip install -r requirements.txt

4. Tesseract OCR (for image and PDF OCR)

   # macOS
   brew install tesseract
   # Ubuntu
   sudo apt-get install tesseract-ocr

5. Poppler (for PDF → image conversion)

   # macOS
   brew install poppler
   # Ubuntu
   sudo apt-get install poppler-utils

6. Google Gemini API Key

   • Generate at Google AI Studio → API Keys

   • Store in .env:

     GOOGLE_API_KEY=AIzaSy…<your_gemini_api_key>…

---

⚙️ Setup & Run

1. Clone repository

   git clone [https://github.com/omar-abdel-aziz/med-qa.git]
   cd medical-qa-api

2. Activate virtual environment

   source venv/bin/activate

3. Install dependencies

   pip install -r requirements.txt

4. Create .env with your GOOGLE_API_KEY

5. Run server

   uvicorn app.main:app --reload --host 127.0.0.1 --port 8000

---

🖥️ API Endpoints

All endpoints are prefixed by / on http://127.0.0.1:8000:

Endpoint	Method	Description
/upload	POST	Upload a PDF/image. Returns session_id.
/process/{session}	POST	Process uploaded file (OCR, chunk, embed, index).
/status/{session}	GET	Check if processing complete. Returns { processed: bool }.
/query/{session}	POST	Ask a question. Returns bullet‑point summary in answer.
/cleanup/{session}	DELETE	Delete session data.

1. Upload

POST /upload
Content-Type: multipart/form-data

Body:
  file: <PDF or image>

Response:
{
  "session_id": "<sid>"
}

2. Process

POST /process/{sid}

Response:
{
  "status": "done"
}

3. Status

GET /status/{sid}

Response:
{
  "processed": true
}

4. Query

POST /query/{sid}
Content-Type: application/json

Body:
{
  "question": "Your medical question"
}

Response:
{
  "answer": [
    "- Bullet point 1",
    "- Bullet point 2",
    "…"
  ]
}

5. Cleanup

DELETE /cleanup/{sid}

Response:
{
  "deleted": true
}

---

🔧 Internal Modules

• app/ocr.py: Extracts text via OCR
• app/pipeline.py: Splits, embeds, and persists FAISS index
• app/main.py: Defines FastAPI routes and RAG workflow

---

🌐 Frontend Example

Use Axios in React to interact:

import axios from "axios";
const api = axios.create({ baseURL: "http://127.0.0.1:8000" });

// Upload
const { data } = await api.post("/upload", formData);
const sid = data.session_id;

// Process
await api.post(`/process/${sid}`);

// Query
const res = await api.post(`/query/${sid}`, { question: "..." });
console.log(res.data.answer);

---

📝 Notes

• Local demo—no external storage or paid services.
• All session data under ./data/{session_id} is removed by /cleanup.
• For production, tighten CORS in app/main.py.

---