12 · Chunking Strategies

The quality of your chunks is the quality of your retrieval. This compares chunking strategies and when to use each, because good chunks are meaningful, self-contained and easy to retrieve.

How it works

1. 1Fixed-size chunking: split by characters with overlap — simple and fast.
2. 2Sentence chunking: split by sentences to keep them intact.
3. 3Paragraph chunking: preserves context for longer form.
4. 4Semantic chunking: split by meaning using embeddings.
5. 5Parent-document: small chunks for retrieval mapped to a larger parent for context.

Key points

• Match the strategy to content: legal docs → paragraph/parent; code → function-based; chat → message-based.
• Typical settings: 500–1200 chars with 100–250 overlap.
• Overlap prevents meaning that spans a boundary from being lost.
• Good chunks are meaningful, self-contained and retrievable.

Why it matters

Chunking is an underrated lever: poor chunks wreck retrieval no matter how good the model is. The right strategy makes RAG precise and grounded.

Frequently asked questions

What is chunk overlap?

Repeating some text between adjacent chunks so ideas spanning a boundary aren't lost.

Which strategy is best?

It depends on content type — semantic or parent-document for context-heavy docs, function-based for code.