UnderstandPDF

A web application that lets users upload PDF documents and instantly see AI-generated insights extracted from them. Instead of reading long documents, users can browse AI-generated cards summarizing key points with source citations while viewing the PDF side by side.

🚀 Features

• PDF Upload - Upload any PDF document for processing
• AI-Generated Insights - Automatically extract key insights as cards with titles, descriptions, and source citations
• Two-Pane Interface - PDF viewer on left, insights sidebar on right
• Interactive Citations - Click page references in insights to navigate PDF

🔧 Key Components

PDF Processing Flow

1. Upload - User uploads PDF via drag-and-drop
2. Processing - Async background processing: text extraction → chunking → embeddings → vector indexing
3. Insights Generation - AI extracts key insights with citations

RAG Architecture

This project implements Retrieval-Augmented Generation (RAG) to provide accurate, citation-backed answers from PDF documents:

1. Document Chunking - PDFs are split into semantic chunks (typically 200-300 words) to balance context and retrieval efficiency
2. Vector Embeddings - Each chunk is converted to a vector embedding using Gemini API
3. Vector Storage - Embeddings stored in Supabase PostgreSQL with pgvector extension
4. Semantic Search - When users ask questions, the system:
   • Generates query embedding
   • Performs vector similarity search (cosine similarity)
   • Retrieves top-K most relevant chunks (typically 5-10)
5. Context Assembly - Retrieved chunks are combined with the user's question
6. LLM Processing - Gemini API generates answers using the assembled context
7. Citation Mapping - The system tracks which chunks were used to generate each answer, providing page-level citations

Benefits of RAG Approach

• Accuracy: Answers are grounded in actual document content
• Cost Efficiency: Only relevant context is sent to the LLM, reducing token usage
• Citation Tracking: Every answer can be traced back to specific document pages
• Scalability: Vector search scales efficiently with document size
• Privacy: Documents never leave the user's account context

Vector Search Implementation

• Similarity Metric: Cosine similarity for semantic matching
• Search Scope: Scoped to individual documents to prevent cross-document leakage
• Performance: Optimized for fast retrieval (sub-second response times)
• Relevance Ranking: Results ordered by similarity score

🚀 Getting Started

1. Clone the repository
2. Set up environment variables using .env.example
3. Install dependencies: npm install
4. Run locally: npm run dev
5. Build for production: npm run build

📄 License

This project is for educational purposes. Please check individual dependencies for their respective licenses.

---

Built with ❤️ using Next.js, Supabase, and Gemini