Function-level or section-level chunking for code modules — Code or Hierarchical Specifications

Tree-aware chunking that follows code or specification hierarchy — Code or Hierarchical Specifications

Per-record chunking where each record is one chunk — Heterogeneous Short Records

Semantic chunking along clause or paragraph boundaries — Long Structured Documents

Fixed-size chunking with overlap for short records of similar length — Heterogeneous Short Records

Hierarchical chunking that mirrors document section structure — Long Structured Documents

Chunking must preserve the structural unit that carries meaning in the source corpus. Code and hierarchical specifications are best divided at function, module, class, or tree boundaries because arbitrary token cuts can separate definitions from their implementation or parent context. Heterogeneous short-record collections should generally preserve each record as an independent chunk. Fixed-size overlapping chunks are also effective when records have broadly similar lengths and lack meaningful internal hierarchy. Long structured documents benefit from semantic boundaries such as clauses and paragraphs, while hierarchical chunking preserves relationships among sections, subsections, and parent headings. Anthropic notes that chunk size, boundaries, and overlap materially affect retrieval performance; therefore, one universal chunking method is inappropriate. Anthropic Contextual Retrieval