👁️

VisionRAG

GitHub ↗

The Business Problem

Standard RAG systems only understand text — they fail on scanned PDFs, invoices, and reports where the visual layout carries meaning.

The Solution

A multimodal document intelligence system that reads the visual layout of each page with GPT-4o Vision before answering questions against it.

The Result

Every answer traces back to the exact page it came from — document and page-level citations, no hallucinated sources
Handles scanned PDFs, invoices, and screenshots, not just clean text documents
Fully containerised via Docker Compose — one command spins up the app and vector database together

Architecture

Document Upload -> Page-Level Vision Embedding (GPT-4o Vision) -> pgvector Similarity Search -> Cited Answer

⚙️ Technical Details For engineers & technical reviewers — click to expand

Every answer traces back to the exact page it came from — document-level and page-level citations, no hallucinated sources. A multimodal document intelligence system that processes scanned PDFs, invoices, and screenshots using GPT-4o Vision. Most RAG systems only handle text; VisionRAG understands the visual layout of each page first, then stores those embeddings in PostgreSQL with pgvector for similarity search. Fully containerised via Docker Compose — one command spins up both the app and the vector database.

GPT-4o Vision pgvector PostgreSQL FastAPI Streamlit Docker pdf2image PyMuPDF

Need something like this for your own business?

Build Something Similar →

Have questions about this project? Ask my AI assistant for details.

Ask AI about this →