📑 FDA RAG Document Intelligence

AI-powered FDA document intelligence system using Retrieval-Augmented Generation (RAG) to analyze FDA 483 inspection reports and provide accurate, compliance-grade answers through a Streamlit interface.

🔍 Project Overview

This project demonstrates an end-to-end RAG (Retrieval-Augmented Generation) pipeline designed to work with FDA inspection and quality documents (FDA 483 reports).

Instead of relying on generic AI knowledge, the system:

• Retrieves relevant content directly from uploaded FDA PDFs
• Uses a strict Quality & Compliance expert prompt
• Generates fact-based, non-hallucinated answers
• Supports querying across all documents or a single selected PDF

The application is built as a professional Streamlit web app, suitable for demos, learning, and portfolio use.

✨ Key Features

• 📄 Query across all FDA PDFs or a single selected document
• 🔍 Accurate document-grounded answers using RAG
• 🧠 Strict FDA Quality, Compliance & R&D expert behavior
• 🚫 No hallucination (answers only from provided PDFs)
• 📊 Displays total PDFs and knowledge chunks
• 🎨 Clean, enterprise-style Streamlit UI
• ⚡ Powered by OpenAI embeddings + Pinecone vector DB

🧠 What is RAG (Retrieval-Augmented Generation)?

RAG combines:

1. Retrieval – finding relevant document chunks from a vector database
2. Generation – using an LLM to answer based only on retrieved content

This ensures:

• High factual accuracy
• No guessing
• Enterprise-ready AI behavior

🏗️ Architecture (High Level)

FDA PDFs ↓ Text Extraction + Chunking ↓ Embeddings (OpenAI) ↓ Vector Database (Pinecone) ↓ Retriever ↓ LLM (Strict Quality Expert Prompt) ↓ Streamlit UI

🧰 Tech Stack

• Language: Python
• Frontend: Streamlit
• LLM & Embeddings: OpenAI
• Vector Database: Pinecone
• RAG Architecture
• PDF Handling: PyMuPDF + OCR (in notebook)

📁 Project Structure

fda-rag-document-intelligence/
│
├── Streamlit_app.py          # Streamlit web application
├── requirements.txt          # Python dependencies
├── README.md                 # Project documentation
├── .gitignore                # Ignored files
│
├── notebooks/
│   └── FDA_BOT_20_01.ipynb   # Original RAG development notebook
│
├── visuals/
│   └── app_screenshots.png   # UI screenshots (optional)
│
└── .streamlit/
    └── secrets.example.toml  # Example secrets file

---

⚙️ Setup Instructions (Run Locally)

1️⃣ Clone the Repository

git clone https://github.com/MK1404/fda-rag-document-intelligence.git
cd fda-rag-document-intelligence

---

2️⃣ Install Dependencies

pip install -r requirements.txt

---

3️⃣ Add API Keys (Secrets)

Create a folder:

mkdir .streamlit

Create a file: .streamlit/secrets.toml

OPENAI_API_KEY = "your-openai-api-key"
PINECONE_API_KEY = "your-pinecone-api-key"

⚠️ Do NOT commit this file to GitHub

---

4️⃣ Run the Streamlit App

streamlit run Streamlit_app.py

---

🖥️ How to Use the App

1. By default, the app searches across all PDFs

2. Use the sidebar dropdown to select a specific PDF

3. Ask questions such as:

   • “List all observations for this site”
   • “Return common FDA observations across all PDFs”
   • “What quality issues were identified?”

4. The system responds using only document content

---

📌 Example Questions

• Return all FDA observations from the selected report
• What repeated quality issues appear across inspections?
• List CAPA-related observations
• Identify compliance gaps mentioned in the document

---

🚀 Future Enhancements

• PDF upload directly from UI
• Source citation highlighting
• Observation severity tagging
• Compliance dashboards
• Export responses as reports

---

⚠️ Important Notes

• This project is for learning and demonstration purposes
• No confidential or proprietary FDA data is included
• Users must provide their own API keys

---

👤 Author

Mohit Data Analytics & AI (Learning Project)

---

⭐ If You Find This Useful

Consider starring ⭐ the repo to support the project.