Summary
Chunking splits documents into pieces short enough to embed. The strategy determines what your retriever can find. Bad chunking is the single most common cause of bad RAG results.
Objectives
- 01State the three main chunking strategies.
- 02Choose a chunk size for a specific corpus.
- 03Explain when to add chunk overlap.
The Lesson
The strategies
Fixed-size (e.g., 512 tokens): simplest, but breaks semantic units. Sentence or paragraph: preserves boundaries at the cost of variable size. Semantic (using embeddings to find break points): best quality, slowest to build. Start with paragraph chunking; move to semantic if quality demands.
Key Ideas
- Chunking determines what you can retrieve.
- Overlap prevents boundary loss. 10-20% is typical.