Ask Book

A RAG (Retrieval-Augmented Generation) tool that lets you ask questions about PDF books using natural language.

How It Works

1. Load a PDF — Text is extracted and split into ~1000-character chunks
2. Store in ChromaDB — Chunks are converted to embeddings and stored in a vector database
3. Ask questions — Your question is matched against the chunks to find relevant passages
4. Get answers — The relevant passages are sent to Claude, which generates an answer

Setup

Python environment

First, make sure you have Python 3.8+ installed. You can check with:

python --version

Create and activate a virtual environment:

macOS/Linux:

python -m venv venv
source venv/bin/activate

Windows (Command Prompt):

python -m venv venv
venv\Scripts\activate.bat

Windows (PowerShell):

python -m venv venv
venv\Scripts\Activate.ps1

Note: On Windows, if you get an execution policy error in PowerShell, run Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser first.

Install dependencies

With your virtual environment activated:

pip install chromadb pymupdf anthropic python-dotenv

Set your API key

Create a .env file in the project directory:

ANTHROPIC_API_KEY=your-key-here

Get a key at https://console.anthropic.com/settings/keys

Usage

Load a book

python ask_book.py load path/to/book.pdf

This only needs to be done once per book. The data is stored in ./chroma_db/.

Ask questions

python ask_book.py ask "Who is the main character?"
python ask_book.py ask "What happens at the end?"
python ask_book.py ask "Describe the relationship between Alice and Bob"

Technology

ChromaDB (Vector Database)

ChromaDB stores text as high-dimensional vectors (embeddings). When you ask a question, it converts your question to a vector and finds the most similar chunks using cosine similarity. This is much more powerful than keyword search — it understands meaning, not just words.

• Runs locally, no external service needed
• Uses all-MiniLM-L6-v2 for embeddings by default
• Data persists in ./chroma_db/

Claude API

Claude (by Anthropic) is the LLM that reads the retrieved passages and generates answers. The app uses claude-sonnet-4-20250514 for a good balance of speed and quality.

The API is pay-per-use. Typical costs for this app:

• ~$0.003 per question (input tokens for context)
• ~$0.015 per answer (output tokens)

Limitations

• Only stores one book at a time (loading a new book replaces the old one)
• Very long books may have less accurate retrieval
• Answers are only as good as the relevant chunks found