📄 DocVision AI: Chat with Your Documents

An intelligent RAG (Retrieval-Augmented Generation) system that lets you upload any PDF and have a conversation with it. Ask questions in natural language, get accurate answers with source citations from your document.

How It Works

Upload PDF → Extract Text → Chunk → Embed → Store in Vector DB → Ask Questions → Get Answers with Sources

Architecture

PDF is loaded using PyMuPDF, split into 1000-character chunks with 200-character overlap using RecursiveCharacterTextSplitter, converted to 1536-dimensional vectors using OpenAI Embeddings, and stored in ChromaDB for fast similarity search. When a user asks a question, the 4 most relevant chunks are retrieved and sent along with the question to GPT-4o-mini, which generates a grounded answer with page-level citations. The entire UI is built with Streamlit.

Features

• Upload any PDF and chat with it instantly
• Answers cite exact pages and passages used
• Persistent vector store — no re-processing on reload
• Clean Streamlit web interface
• Source transparency — see which chunks were used for each answer

Tech Stack

• Orchestration: LangChain (LCEL — LangChain Expression Language)
• LLM: OpenAI GPT-4o-mini
• Embeddings: OpenAI text-embedding-ada-002
• Vector Database: ChromaDB (local)
• PDF Processing: PyMuPDF (fitz)
• Frontend: Streamlit
• Language: Python 3.10+

Quick Start

1. Clone the repo

git clone https://github.com/VdKPr/DocVision-AI.git
cd DocVision-AI

2. Create virtual environment

python -m venv venv
venv\Scripts\activate

3. Install dependencies

pip install -r requirements.txt

4. Set up API key

Create a .env file in the project root:

OPENAI_API_KEY=your-openai-api-key-here

Get your key from: https://platform.openai.com/api-keys

5. Run the app

streamlit run docvision.py

App opens at http://localhost:8501

Project Structure

DocVision-AI/
├── docvision.py          # Main application
├── requirements.txt      # Python dependencies
├── .env                  # API key (not tracked by git)
├── .gitignore            # Git ignore rules
└── README.md             # This file

How RAG Works (for the curious)

1. Load — PyMuPDF extracts text from each page of the uploaded PDF
2. Chunk — Text is split into 1000-character pieces with 200-character overlap to preserve context across boundaries
3. Embed — OpenAI Embeddings converts each chunk into a 1536-dimensional vector that captures semantic meaning
4. Store — ChromaDB indexes these vectors locally for fast cosine similarity search
5. Retrieve — When you ask a question, it gets embedded too, and the 4 most similar chunks are found
6. Generate — GPT-4o-mini receives your question + the retrieved chunks and produces a grounded answer with citations

Roadmap

• Multimodal RAG — understand images, charts, and diagrams inside PDFs using CLIP + GPT-4o vision
• Multi-document support — upload and query across multiple PDFs simultaneously
• Chat history — maintain conversation context across questions
• Local LLM support via Ollama (Llama 3) for fully offline, privacy-preserving usage
• Evaluation metrics — RAGAS-based retrieval and generation quality scoring
• Deploy to Streamlit Cloud with public demo URL

Built With

This project was built as part of my transition into Multimodal AI Engineering, combining my background in physics-informed machine learning (M.Tech thesis at IIT Bombay) with modern GenAI tools and frameworks.