RAG Patterns Compared: KG-RAG vs Hybrid RAG vs Agentic RAG

An educational repo that builds the same Q&A system three ways so you can see, run, and measure the differences.

Patterns covered

Pattern	Core idea	Strongest when
KG-RAG	Retrieve from a knowledge graph; answer with structured facts and relationships	Multi-hop reasoning, entity-rich domains (biomed, finance, legal)
Hybrid RAG	Fuse dense (embeddings) + sparse (BM25) retrieval, then rerank	Mixed terminology, exact terms matter, broad corpora
Agentic RAG	An agent decides which tool/index to query, refines, and retries	Heterogeneous sources, complex queries, tool use needed

Learning path

1. Read docs/01-concepts.md
2. Run notebooks/00_setup_and_data.ipynb to load the shared corpus
3. Walk through patterns one-by-one: notebooks 01, 02, 03
4. See them on the same questions in notebooks/04_side_by_side_comparison.ipynb
5. Measure quality, latency, and cost in notebooks/05_evaluation_and_costs.ipynb

Quickstart

# 1. Install (uses uv; falls back to pip)
uv sync   # or: pip install -e ".[dev]"

# 2. Copy env template and add API keys
cp .env.example .env

# 3. Start backing services (Neo4j + Qdrant)
docker compose up -d

# 4. Launch notebooks
jupyter lab

Repo layout

data/        Source corpus + processed artifacts
docs/        Concepts, architecture diagrams, tradeoffs, decision guide
src/         Python package implementing each pattern + shared utilities
notebooks/   Hands-on walkthroughs and the side-by-side comparison
tests/       Smoke tests so notebooks don't rot

When to use which pattern

See docs/04-when-to-use.md for a decision tree. Short version:

• Need multi-hop joins across entities? → KG-RAG
• Need best general retrieval quality with low effort? → Hybrid RAG
• Need to route across multiple sources or refine queries? → Agentic RAG

License

MIT. Corpus licenses noted in data/README.md.