rag-playbook 0.2.2

Compare RAG patterns with real benchmarks. Run one command, see which pattern wins.

rag-playbook

Stop guessing which RAG pattern to use. Compare them with real numbers.

Quick Start · Patterns · Decision Guide · Architecture · CLI Reference

---

Every RAG tutorial teaches you how to build patterns. None of them tell you which one to actually use.

rag-playbook runs the same query against 8 production-tested RAG patterns and shows you which one wins — with real numbers for quality, latency, and cost.

Quick Start

pip install rag-playbook[openai]
export OPENAI_API_KEY=sk-...

# Compare all patterns on your documents
rag-playbook compare --data ./my_docs/ --query "What is the refund policy?"

Patterns

#	Pattern	Best For	Latency	Cost
01	Naive	Simple factual queries	~1s	$
02	Hybrid Search	Queries with codes, IDs, exact terms	~1.1s	$
03	Re-ranking	When top-K retrieval isn't precise enough	~1.4s	$$
04	Parent-Child	Long documents with clear sections	~1s	$
05	Query Decomposition	Complex multi-part questions	~2.1s	$$$
06	HyDE	Short or ambiguous queries	~1.5s	$$
07	Self-Correcting	When hallucination risk is high	~2.8s	$$$
08	Agentic	When query intent is unclear	~3.2s	$$$$

Which pattern should I use? (decision guide)

How is this different from [X]?

	NirDiamant/RAG_Techniques	FlashRAG	Ragas	rag-playbook
Format	Jupyter notebooks	Academic library	Eval metrics	Library + CLI
pip install	No	Complex	Yes	Yes (simple)
Benchmarks	None	Academic	N/A	Practical comparison
"Which to use?"	No	No	No	YES
License	Non-commercial	MIT	Apache-2.0	MIT

Use as a Library

import asyncio
from rag_playbook import create_pattern, Settings
from rag_playbook.core.embedder import create_embedder
from rag_playbook.core.llm import create_llm
from rag_playbook.core.vector_store import create_vector_store

async def main():
    settings = Settings()  # Reads from .env / environment
    llm = create_llm(settings)
    embedder = create_embedder(settings)
    store = create_vector_store("memory")

    pattern = create_pattern("reranking", llm=llm, embedder=embedder, store=store)

    # Index your documents first (or use the CLI: rag-playbook ingest)
    result = await pattern.query("What is the refund policy?")

    print(result.answer)
    print(f"Cost: ${result.metadata.cost_usd:.4f}")
    print(f"Latency: {result.metadata.latency_ms:.0f}ms")

asyncio.run(main())

See examples/ for more usage patterns.

Installation

# Minimal (in-memory store, OpenAI)
pip install rag-playbook[openai]

# Everything (includes planned vector store backends)
pip install rag-playbook[all]

From Source

git clone https://github.com/Aamirofficiall/rag-playbook.git
cd rag-playbook
uv venv --python 3.11 .venv
source .venv/bin/activate
uv pip install -e ".[dev,chromadb,openai]"

CLI Commands

Command	Description
rag-playbook compare	Compare patterns side-by-side on your documents
rag-playbook run	Run a single pattern
rag-playbook recommend	Get an LLM-powered pattern recommendation
rag-playbook ingest	Load, chunk, embed, and index documents
rag-playbook bench	Run full benchmark suite
rag-playbook patterns	List all available patterns

Run a single pattern

rag-playbook run reranking --data ./docs/ --query "What are the laptop specs?"

List available patterns

rag-playbook patterns

Get a pattern recommendation

rag-playbook recommend --query "What is the refund policy?"

Ingest documents

rag-playbook ingest --data ./sample_docs/

See CLI Reference for full usage.

Configuration

Environment variables map directly to settings — no prefix needed:

# .env — Core settings
OPENAI_API_KEY=sk-...                  # Required for OpenAI provider
DEFAULT_LLM_PROVIDER=openai            # openai | anthropic
DEFAULT_LLM_MODEL=gpt-4o-mini         # Any model your provider supports
EMBEDDING_MODEL=text-embedding-3-small
EMBEDDING_DIMENSION=1536
DEFAULT_TOP_K=5
DEFAULT_CHUNK_SIZE=512
DEFAULT_CHUNK_OVERLAP=50

Using OpenRouter (or any OpenAI-compatible API)

Works with any OpenAI-compatible endpoint — OpenRouter, Azure OpenAI, Ollama, vLLM, etc:

OPENAI_API_KEY=sk-or-v1-...                     # Your OpenRouter key
OPENAI_BASE_URL=https://openrouter.ai/api/v1    # Custom base URL
DEFAULT_LLM_MODEL=openai/gpt-4o-mini            # OpenRouter model format

Vector store backends

The default in-memory store works out of the box but re-embeds every session. ChromaDB, pgvector, and Qdrant are planned as persistent backends — the optional dependencies are installable but the store implementations are not yet wired up. Contributions welcome.

Or pass a Settings object directly in code. See .env.example for all options.

Architecture

Document → Chunk → EmbeddedChunk → RetrievedChunk → RAGResult
              │         │                │               │
          Chunker    Embedder       VectorStore        LLM

LLM providers, embedders, and vector stores are all swappable. Each RAG pattern extends a base class and overrides only the steps it needs. Embeddings are cached automatically so running compare across all 8 patterns doesn't re-embed your docs 8 times.

See Architecture Guide for details.

Development

make install    # Install with dev dependencies
make test       # Run unit tests
make lint       # Lint with ruff
make format     # Auto-format with ruff
make type-check # Type check with mypy
make check      # Run all checks

See CONTRIBUTING.md for the full guide.

Author

Built by Aamir Shahzad — backend engineer building data systems and AI infrastructure.

License

MIT