# Chunking Strategy

How you split source documents into retrievable units — it sets the ceiling on everything retrieval can return.

## Why the CCAR-P exam tests this

Chunking is the most common lever for fixing bad RAG results, so exams probe whether you understand the size/overlap/structure trade-offs.

## Definition

**Chunking** divides documents into passages that get embedded and retrieved independently. **Too large** and each chunk carries diluted, multi-topic embeddings plus irrelevant filler that wastes context; **too small** and chunks lose the surrounding context needed to be meaningful. **Overlap** (repeating a few sentences between adjacent chunks) preserves continuity across boundaries.

Strategies range from **fixed-size** (by tokens/characters) to **recursive/structural** splitting that respects natural boundaries — headings, paragraphs, sentences, code blocks, or table rows. **Semantic chunking** groups sentences by embedding similarity so each chunk is one coherent idea. Structure-aware splitting usually beats naive fixed-size on real documents.

Attach **metadata** to each chunk (source, section, `doc_id`, timestamps, and crucially **ACL/owner fields**) so retrieval can filter and so downstream authorization can be enforced. Chunk boundaries also carry the identifiers used for per-document access control at retrieval time.

## Exam trap

There is no universal chunk size — the right size depends on document structure and query type, so an answer asserting a fixed 'best' token count is a trap.

## Commonly confused with

- **reranking** — Chunking decides what units exist to retrieve; reranking reorders whatever was retrieved. Different stages.
- **rag-pipeline** — Chunking is the indexing step within the broader RAG pipeline.

## Related

- rag-pipeline
- reranking
- hybrid-retrieval
- data-layer-acls

---

Source: Claude Cert Prep — an independent, unofficial CCAR-P study resource (domain P3). Not affiliated with Anthropic.