The RAG Cookbook 2026

Welcome. This site renders every recipe in this repository as a readable page, with the source notebooks just a click away.

If you are new here, start with the cookbook tour and then the vanilla pipeline. After that, the Theory section gives the conceptual map and the Recipes section is the cookbook itself.

What you will learn

• How to chunk documents in ways that beat naive fixed-window splitting (recipes 5–12)
• How to rewrite queries so retrieval finds what the user actually meant (recipes 13–17)
• How to combine dense, sparse, and late-interaction retrievers (recipes 18–21)
• How to rerank top-k with cross-encoders and LLMs (recipes 22–23)
• How to build adaptive and agentic pipelines that decide for themselves whether and how to retrieve (recipes 24–30)
• How to ground retrieval in knowledge graphs and long-term memory (recipes 31–34)
• How to retrieve from PDFs as images, no OCR (recipes 35–36)
• How to evaluate, trace, and harden a pipeline for production (recipes 37–40)

Conventions

• The default provider is Nebius AI Studio. Every recipe begins with one cell that lets you flip to OpenAI, Anthropic, Groq, OpenRouter, or Together.
• The default vector store is Qdrant in-memory. Recipes that benchmark stores use Qdrant, LanceDB, and Chroma side by side.
• All recipes use the four-corpus mix described on the Recipes index.

Happy retrieving.