PDF AI Assistant

A full-stack AI application that allows users to upload a PDF, generate a summary, and ask questions about its content.

The application uses Retrieval-Augmented Generation (RAG) to retrieve relevant sections from the PDF before generating answers. This helps ensure responses are grounded in the uploaded document and include page references.

Architecture

flowchart LR
    User[User]

    subgraph Frontend["Next.js Frontend"]
        Upload[PDF Upload]
        Summary[Document Summary]
        Chat[CopilotKit Chat]
        Runtime[Copilot Runtime]
    end

    subgraph Backend["FastAPI Backend"]
        API[Document API]
        Processor[PDF Processing]
        Agent[LangGraph Agent]
        Search[Document Search Tool]
    end

    OpenAI[OpenAI]
    Pinecone[(Pinecone)]

    User --> Upload
    Upload --> API
    API --> Processor

    Processor --> OpenAI
    Processor --> Pinecone
    Processor --> Summary

    User --> Chat
    Chat --> Runtime
    Runtime --> Agent
    Agent --> Search
    Search --> Pinecone
    Agent --> OpenAI
    Agent --> Chat

Application Workflow

PDF processing

flowchart TD
    A[Upload PDF] --> B[Validate file]
    B --> C[Extract text page by page]
    C --> D[Split text into chunks]
    D --> E[Generate embeddings]
    E --> F[Store vectors in Pinecone]
    F --> G[Generate document summary]
    G --> H[Delete temporary PDF]
    H --> I[Display summary]

Question answering

flowchart TD
    A[User asks a question] --> B[CopilotKit sends request]
    B --> C[LangGraph agent receives question]
    C --> D[Search relevant chunks in Pinecone]
    D --> E[Return text and page metadata]
    E --> F[OpenAI generates grounded answer]
    F --> G[Stream answer with page citations]

How RAG Works

The application does not send the complete PDF to the language model for every question.

Instead, it:

1. Extracts the PDF text.
2. Splits the text into smaller chunks.
3. Converts each chunk into an embedding.
4. Stores the embeddings in Pinecone.
5. Searches for chunks related to the user's question.
6. Sends only the relevant chunks to the language model.
7. Generates an answer with page citations.

Features

• Upload text-based PDF files
• Extract and split PDF content into chunks
• Generate and store embeddings in Pinecone
• Automatically summarize uploaded documents
• Ask questions through an AI chat interface
• Retrieve relevant document sections using semantic search
• Return answers with page citations
• Delete document vectors after removing a PDF
• Support switching from OpenAI to Ollama later

Tech Stack

Frontend

• Next.js App Router
• TypeScript
• React
• Tailwind CSS
• CopilotKit
• AG-UI Client

Backend

• Python
• FastAPI
• LangChain
• LangGraph
• OpenAI
• Pinecone
• PyPDF
• AG-UI LangGraph

Project Structure

pdf-ai-assistant/
├── backend/
│   ├── app/
│   │   ├── agents/
│   │   ├── api/
│   │   ├── core/
│   │   ├── models/
│   │   ├── services/
│   │   └── main.py
│   ├── tests/
│   └── pyproject.toml
│
├── frontend/
│   ├── src/
│   │   ├── app/
│   │   ├── components/
│   │   └── lib/
│   └── package.json
│
└── README.md

Getting Started

Requirements

• Python 3.12+
• Node.js 20+
• uv
• pnpm
• OpenAI API key
• Pinecone API key

Backend environment

Create backend/.env:

OPENAI_API_KEY=your-openai-api-key

PINECONE_API_KEY=your-pinecone-api-key
PINECONE_INDEX_NAME=pdf-ai-assistant
PINECONE_CLOUD=aws
PINECONE_REGION=us-east-1

CHAT_MODEL=gpt-4.1-mini
EMBEDDING_MODEL=text-embedding-3-small

MAX_UPLOAD_SIZE_MB=20
FRONTEND_ORIGIN=http://localhost:3000

Frontend environment

Create frontend/.env.local:

BACKEND_URL=http://127.0.0.1:8000
NEXT_PUBLIC_BACKEND_URL=http://localhost:8000

Start the backend

cd backend
uv sync
uv run uvicorn app.main:app --reload --port 8000

Backend API:

http://localhost:8000

API documentation:

http://localhost:8000/docs

Start the frontend

cd frontend
pnpm install
pnpm dev

Frontend:

http://localhost:3000

License

This project is intended for learning and experimentation.