Mastering RAG

The One Chunking Book That Actually Ran the Benchmarks

Your RAG accuracy is stuck at 62%. You've tried better embeddings, bigger models, prompt tricks. Nothing moved the needle. The culprit? Chunking - the least glamorous, most impactful part of RAG. This is the deep-dive book on the ONE thing that decided +26% Recall in the author's production system.

Because the "just use RecursiveCharacterTextSplitter(1000, 200)" advice is why your retrieval breaks on real documents. This volume, Book 1 of the Mastering RAG series, is 21 chapters and a real ~50,000-chunk benchmark dedicated to answering one question: which chunking method actually wins, and why?

• Fixed-size: the baseline everyone starts with - and its blind spots
• Recursive character: LangChain's default, its hidden trade-offs
• Sentence & semantic: KSS, spaCy, KoSimCSE percentile splitting
• Small-to-Big & propositional: retrieve fine, return coarse
• Contextual Retrieval (Anthropic): full implementation with prompt caching
• Agentic chunking: LLM as editor, when it's worth the cost
• Late Chunking (Jina): reversing the order for long-context embeddings
• Hybrid strategies: which combinations actually win

• ~50,000 chunks from a real production project (not toy Wikipedia data)
• 15-query golden set: facts, procedures, comparisons, overviews
• Recall@1/5/10, MRR, per-query-type breakdowns
• Cost analysis: Contextual Retrieval at $47 total for +26% Recall
• Why Semantic Chunking underperformed in practice - with root cause
• Baseline reproducibility issues you won't see in blog posts

• Pre-BERT IR era: why chunking didn't exist
• BERT's 512-token ceiling: how chunking became mandatory
• LangChain's standardization era and its lock-in effects
• Semantic chunking emergence, then LLM-based chunking

• Engineers whose RAG plateaued and don't know why
• Tech leads deciding between chunking libraries and strategies
• ML engineers who need to justify chunking choices with numbers
• Anyone who read "Advanced RAG" and wants the full chunking story

Every chunking blog post says "it depends." This book shows you what it depends on - with a benchmark, a golden set, and a $47 invoice. Python and Java implementations included, along with a production deployment checklist and appendix cheat sheets covering every major chunking library.

Volume 1: Chunking Deep Dive (this book) · Volume 2: Retrieval Deep Dive (BM25, hybrid, rerank) · Volume 3: Embeddings and Production Stack. Each volume goes as deep on one topic as most books go across all of RAG.

Basic RAG knowledge (or read "AI RAG for Web Developers" first). Comfortable with Python; Java examples optional.

Stop guessing at chunk sizes. Read the benchmark, ship the answer.