PDF Q&A Chatbot

A document question-answering assistant that takes a PDF, extracts text, splits it into chunks, generates embeddings, and answers user questions using retrieved context.

Built with Google Gemini (via OpenAI-compatible API) and NumPy. No LangChain, no vector databases.

Setup

python -m venv .venv
.venv\Scripts\activate        # Windows
# source .venv/bin/activate   # macOS/Linux

pip install -r requirements.txt

Create a .env file with your Gemini API key:

GOOGLE_API_KEY=your_key_here

Get one at https://aistudio.google.com/apikey

Usage

# Run the Streamlit Web UI
streamlit run app.py

# Or use the CLI
python main.py path/to/document.pdf

Options:

• --chunk-size N — characters per chunk (default 500)
• --overlap N — overlap between chunks (default 50)
• --top-k N — number of chunks to retrieve (default 3)
• --show-chunks — show which chunks were retrieved

Project Structure

pdf_loader.py    — extract text from PDF pages
chunker.py       — split text into overlapping chunks
embeddings.py    — generate embeddings via Gemini API
retriever.py     — cosine similarity search with NumPy
llm_client.py    — send context + question to Gemini LLM
app.py           — Streamlit Web UI entry point
main.py          — CLI entry point, ties everything together

How It Works

1. Load PDF and extract text from all pages using pypdf
2. Split the text into ~500 character chunks with 50 char overlap
3. Generate embeddings for all chunks using gemini-embedding-001
4. For each question:
   • Embed the question
   • Compute cosine similarity (NumPy) against all chunk embeddings
   • Take top-3 most similar chunks
   • Send those chunks + question to gemini-2.5-flash
   • Print the answer
5. Conversation history (last 3 turns) is included for follow-up questions