The RAG Cookbook 2026

Recipes for production-grade retrieval — forty hands-on chapters covering the state of the art in retrieval-augmented generation, all runnable end-to-end on Nebius AI Studio (or any OpenAI-compatible provider).

---

Why this exists

Most RAG tutorials stop at "embed a PDF, do cosine search, hand it to GPT." That worked in 2023. In 2026, production RAG looks completely different: late chunking, contextual retrieval, multi-vector late interaction, listwise rerankers, speculative drafters, agentic workflows over MCP servers, dual-level graph indexes, page-as-image vision retrieval, and compiled DSPy pipelines that out-perform anything a prompt engineer can hand-tune.

This cookbook teaches all of it. Forty self-contained recipes. Every recipe ships with theory, implementation, output discussion, and an evaluation harness. Nothing is hidden behind framework magic.

Quick start

# 1. Clone and install
git clone https://github.com/FareedKhan-dev/rag-cookbook-2026.git
cd rag-cookbook-2026
uv sync                          # or: pip install -e ".[docs,dev]"

# 2. Add your keys (Nebius is the default; OpenAI/Anthropic/Groq optional)
cp .env.example .env
$EDITOR .env

# 3. Download the public-domain corpus (one-time, ~150 MB)
uv run python scripts/fetch_corpus.py

# 4. Launch a recipe
uv run jupyter lab recipes/01-foundations/vanilla-pipeline.ipynb

Want to read the cookbook in your browser instead? uv run mkdocs serve → http://localhost:8000

Recipe Index

#	Category	Recipe	Idea	Best for
1	Foundations	cookbook-tour	Provider switcher, corpus, tracing	First read
2	Foundations	vanilla-pipeline	The baseline you'll beat 38 times	Mental model
3	Foundations	embedding-zoo	Voyage / BGE-M3 / Nebius / Qwen3 head-to-head	Picking a model
4	Foundations	chunk-size-sensitivity	Sweep sizes, measure recall	Tuning
5	Chunking	semantic-boundary-splitting	Split at embedding-distance breakpoints	Mixed-topic docs
6	Chunking	late-chunking-jina	Embed first, chunk after	Long context
7	Chunking	contextual-retrieval-anthropic	LLM-written "where this fits" headers	Highest ROI tweak
8	Chunking	proposition-decomposition	Atomic-claim units	Dense factual recall
9	Chunking	sentence-window-and-parent-child	Small to find, big to read	Precision + synthesis
10	Chunking	raptor-trees	Recursive summary tree	Multi-hop reasoning
11	Chunking	document-summary-routing	Doc summaries → sub-chunks	Heterogeneous corpora
12	Chunking	matryoshka-coarse-to-fine	MRL embeddings, two-stage	Low-latency search
13	Query	hypothetical-document-embeddings	HyDE	Zero-shot retrieval
14	Query	multi-query-rag-fusion	N rewrites + RRF	Recall boost
15	Query	step-back-abstraction	Abstract then ground	Reasoning questions
16	Query	sub-question-decomposition	Tree of sub-queries	Complex queries
17	Query	semantic-router	Classify and dispatch	Multi-index systems
18	Retrieval	hybrid-dense-plus-bm25	Dense + BM25 + RRF on Qdrant	Real workloads
19	Retrieval	colbert-late-interaction	Token-level MaxSim	Highest accuracy
20	Retrieval	auto-retrieval-with-metadata	LLM extracts filters	Structured corpora
21	Retrieval	mmr-diversity	Penalize redundancy	Diverse top-k
22	Reranking	cross-encoder-rerank	bge / Cohere / Jina	Quality lift
23	Reranking	listwise-llm-rerank	RankGPT-style	When budget allows
24	Agentic	self-reflective-retrieval	Self-RAG reflection tokens	Selective retrieval
25	Agentic	corrective-retrieval-with-web-fallback	CRAG	Robust answers
26	Agentic	adaptive-routing-by-question-class	Adaptive RAG	Mixed query types
27	Agentic	speculative-rag-drafter-verifier	Small drafter, big verifier	Latency-bound apps
28	Agentic	langgraph-agentic-rag	Stateful cyclic graph	Tool-using agents
29	Agentic	mcp-tool-retrieval	Retrieval over MCP servers	Modular tools
30	Agentic	dspy-compiled-rag	Optimize, don't prompt-engineer	Repeatable pipelines
31	Graph	microsoft-graphrag-pipeline	Communities + global/local	Global questions
32	Graph	lightrag-dual-level	LightRAG	Cheap incremental KG
33	Graph	hipporag-pagerank-memory	PageRank over KG	Multi-hop facts
34	Graph	mem0-long-term-memory	Persistent agent memory	Chat assistants
35	Multimodal	colpali-page-as-image	VLM patch embeddings of PDFs	No-OCR pipelines
36	Multimodal	vlm-synthesis-over-pages	ColPali → Qwen-VL	Visual answers
37	Eval	ragas-triad-eval	Faithfulness / relevance / precision	Default metrics
38	Eval	deepeval-pytest-ci	Metrics in CI	Regression catch
39	Eval	phoenix-tracing-debugging	End-to-end OTel traces	Debugging
40	Eval	hallucination-guardrails-lynx	Patronus Lynx	Production safety

Stack

Layer	Default	Notes
LLM provider	Nebius AI Studio	Switchable to OpenAI, Anthropic, Groq, OpenRouter, Together via one env var
Provider abstraction	LiteLLM	Single LLMClient wraps chat / embed / rerank
Embeddings	voyage-3-large, BGE-M3, Qwen3-Embedding-8B, Nebius	Compared head-to-head in recipe 3
Rerankers	Cohere Rerank 4, bge-reranker-v2-m3, jina-reranker-v3	Recipe 22
Vector store	Qdrant in-memory by default; LanceDB and Chroma also wrapped	Native hybrid + multi-vector
Graph store	NetworkX (Neo4j optional)	Recipes 31–34
Agent frameworks	LangGraph, LlamaIndex Workflows, DSPy	One recipe each
Tracing	Arize Phoenix (local) + LangSmith (agentic)	One-line init_tracing()
Evaluation	RAGAS + DeepEval	Plus Patronus Lynx guardrails
Documentation	MkDocs Material + mkdocs-jupyter	Served at mkdocs serve

Corpus

Four fresh, public-domain document sets — chosen to exercise every technique without overlapping any other RAG tutorial:

1. A recent arXiv paper on state-space models (not "Attention Is All You Need")
2. A Wikipedia subset on superconductors (~50 markdown pages)
3. Palantir's 2024 SEC 10-K (text extracted, public filing)
4. A curated chapter set from The Rust Programming Language (CC-BY)

Fetch reproducibly with uv run python scripts/fetch_corpus.py --verify.

Notebook structure

Every recipe follows the same six sections:

1. What & Why — original paper / blog, when this technique helps, when it doesn't
2. Setup — provider switch, tracing, corpus load (one cell)
3. Build the Pipeline — the implementation, broken into small reviewable cells
4. Run on Corpus — execute against one of the four corpora
5. Discuss the Output — markdown analysis of what came out, failure modes, observations
6. Evaluate — RAGAS / DeepEval metrics plus one custom assertion

Contributing

Issues and pull requests welcome. New techniques are added under the appropriate category as <technique-slug>.ipynb, following the six-section template above. See docs/theory/ for the conceptual framing you should match.

Citing

If this cookbook helps your research or teaching, please cite it via the CITATION.cff file.

License

Apache-2.0 © 2026 Fareed Khan.