readanybook 0.1.15

A RAG-based cheat sheet generator for books and papers

ReadAnyBook 📚

A RAG-based cheat sheet generator that transforms books and papers into structured, 12-page LaTeX cheat sheets.

🚀 Try It Now

Features

• Multi-format Document Support: PDF, EPUB, HTML, LaTeX, Markdown
• Intelligent Chunking: Math-aware and code-aware text splitting
• Hybrid Retrieval: Dense embeddings + BM25 with reciprocal rank fusion
• Multi-pass Generation: Separate extraction for concepts, formulas, algorithms, and models
• LaTeX Output: Professional cheat sheets compiled to PDF
• Multiple LLM Backends: HuggingFace, Ollama, vLLM, OpenAI-compatible APIs
• Vector Store Options: ChromaDB, Qdrant, Weaviate

Quick Start

Installation

# Basic installation
pip install readanybook

# With CLI support
pip install readanybook[cli]

# With all features
pip install readanybook[all]

From Source

git clone https://github.com/readanybook/readanybook.git
cd readanybook
pip install -e ".[dev]"

Usage

Command Line

# Generate a cheat sheet from a PDF
read-any-book build document.pdf -o cheatsheet.pdf

# Use a specific profile
read-any-book build document.pdf --profile math_paper

# Index a document
read-any-book index document.pdf --collection my_collection

# Search indexed documents
read-any-book search "gradient descent" --collection my_collection

Python API

from readanybook import CheatSheetPipeline, Settings

# Initialize pipeline
settings = Settings()
pipeline = CheatSheetPipeline(settings)

# Process document
pipeline.ingest("textbook.pdf")
pipeline.index(collection_name="textbook")

# Generate cheat sheet
content = pipeline.generate_content()
cheat_sheet = pipeline.build(content, "output/cheatsheet.pdf")

print(f"Generated: {cheat_sheet.pdf_path}")

Quick Start (One-Liner)

from readanybook import build_cheatsheet

# Generate a cheat sheet with a single function call
output = build_cheatsheet(
    "textbook.pdf",
    llm_backend="huggingface",      # or "ollama" for local
    llm_model="Qwen/Qwen2.5-1.5B-Instruct",
    output_format="markdown",        # "latex", "markdown", or "both"
    in_memory=True                   # Required for Kaggle/Colab
)

REST API

# Start the API server
uvicorn readanybook.api:app --host 0.0.0.0 --port 8000

# Upload a document
curl -X POST "http://localhost:8000/upload" \
  -F "file=@document.pdf" \
  -F "collection_name=my_docs"

# Generate cheat sheet
curl -X POST "http://localhost:8000/generate" \
  -H "Content-Type: application/json" \
  -d '{"collection_name": "my_docs", "title": "My Cheat Sheet"}'

Configuration

Create a config.yaml file or use environment variables:

# Embedding model
embedding:
  model_name: "BAAI/bge-base-en-v1.5"
  device: "cuda"

# Vector store
vectordb:
  store_type: "chroma"
  persist_directory: "./data/chroma"

# LLM settings
llm:
  backend: "ollama"
  model_name: "llama3:8b"

# Retrieval
retrieval:
  mode: "hybrid"
  top_k: 15
  
# LaTeX output
latex:
  columns: 2
  font_size: 10
  paper_size: "a4paper"

Configuration Profiles

Use built-in profiles for different document types:

# For technical books
read-any-book build book.pdf --profile technical_book

# For math papers
read-any-book build paper.pdf --profile math_paper

# For non-technical books
read-any-book build novel.pdf --profile nontechnical_book

Architecture

ReadAnyBook follows a modular pipeline architecture with clear separation between ingestion, retrieval, and generation layers.

System Design

┌─────────────┐    ┌─────────────┐    ┌─────────────┐    ┌─────────────┐    ┌─────────────┐    ┌─────────────┐
│  Ingestion  │───▶│  Chunking   │───▶│  Indexing   │───▶│  Retrieval  │───▶│ Generation  │───▶│   LaTeX     │
│  (PDF/EPUB) │    │  (Math-     │    │ (Embeddings)│    │  (Hybrid    │    │  (LLM +     │    │  (Compile   │
│             │    │   aware)    │    │             │    │   Search)   │    │   RAG)      │    │   to PDF)   │
└─────────────┘    └─────────────┘    └──────┬──────┘    └──────┬──────┘    └──────┬──────┘    └─────────────┘
                                             │                  │                  │
                                             ▼                  ▼                  ▼
                                      ┌─────────────┐    ┌─────────────┐    ┌─────────────┐
                                      │ Vector DB   │    │ BM25 Index  │    │ LLM Backend │
                                      │ (ChromaDB)  │    │             │    │  (Ollama/   │
                                      │             │    │             │    │   HF/vLLM)  │
                                      └─────────────┘    └─────────────┘    └─────────────┘

Key Components

Component	Description
Ingestion	Multi-format document parsing (PDF, EPUB, HTML, LaTeX)
Chunking	Hierarchical, semantic, or fixed-size with math/code awareness
Indexing	BGE/E5 embeddings stored in ChromaDB/Qdrant/Weaviate
Retrieval	Hybrid dense+sparse search with RRF fusion and cross-encoder reranking
Generation	Multi-pass extraction: concepts, formulas, algorithms, models
Output	Jinja2 LaTeX templates compiled to 12-page PDF

Package Structure

readanybook/
├── core/           # Domain logic
│   ├── ingestion.py    # Document parsing
│   ├── chunking.py     # Text splitting
│   ├── indexing.py     # Embedding & indexing
│   ├── retrieval.py    # Hybrid retrieval
│   ├── models.py       # LLM clients
│   ├── prompts.py      # Prompt templates
│   └── pipeline.py     # Main orchestrator
├── generation/     # Content generation
│   ├── concepts.py     # Concept extraction
│   ├── formulas.py     # Formula extraction
│   ├── algorithms.py   # Algorithm synthesis
│   ├── models_theory.py # Model summarization
│   └── latex_builder.py # LaTeX generation
├── evaluation/     # Quality metrics
│   ├── rag_eval.py     # RAG evaluation
│   └── metrics.py      # Content metrics
├── infra/          # Infrastructure
│   ├── settings.py     # Configuration
│   ├── vectordb.py     # Vector stores
│   ├── logging.py      # Logging
│   └── tracing.py      # Observability
├── api/            # REST API
├── cli/            # Command line interface
├── templates/      # LaTeX templates
└── config/         # Default configs

Design Principles

• Hexagonal Architecture: Domain services isolated from external adapters
• Configuration-Driven: All behavior controlled via Pydantic settings
• Pluggable Backends: LLM, vector store, and embedding model abstractions
• Observability: Structured logging and tracing throughout

📄 Full documentation: See docs/architecture.pdf for the complete software architecture document.

Requirements

• Python 3.10+
• PyTorch 2.0+
• LaTeX distribution (for PDF compilation)
  • TeX Live, MiKTeX, or Tectonic

LaTeX Installation

# Ubuntu/Debian
sudo apt install texlive-full

# macOS
brew install --cask mactex

# Or use Tectonic (lightweight)
cargo install tectonic

Development

# Install dev dependencies
pip install -e ".[dev]"

# Run tests
pytest

# Format code
black readanybook tests
isort readanybook tests

# Type check
mypy readanybook

# Lint
ruff check readanybook

📓 Kaggle Experiment

We tested ReadAnyBook on Kaggle with Hull's "Options, Futures, and Other Derivatives" (902 pages) using a T4 GPU:

Model	Backend	Time	Notes
microsoft/phi-2	HuggingFace	~15 min	Fast, basic quality
Qwen/Qwen2.5-1.5B-Instruct	HuggingFace	~20 min	Better reasoning
meta-llama/Llama-3.2-3B-Instruct	HuggingFace	~30 min	Best quality (needs HF token)

Note: Use latex_only=True on Kaggle/Colab since pdflatex is not available.

Examples

See the examples directory for:

• 📓 readanybook_demo.ipynb - Interactive notebook tutorial
• Processing academic papers
• Creating ML textbook cheat sheets
• Custom template usage
• API integration examples

License

MIT License - see LICENSE for details.

Contributing

Contributions welcome! Please read CONTRIBUTING.md first.

Acknowledgments

• Built with 🤗 Transformers, ChromaDB, and FastAPI
• Inspired by the need for better study materials