Structural Chunking

The algorithm parses the document's structure to identify heading levels (H1-H6), table boundaries, and list blocks as primary split candidates. Unlike naive methods, it never breaks content mid-paragraph or mid-sentence if a structural marker is available. This maintains the logical flow of information, ensuring that a chunk containing an H2 heading also includes its subordinate paragraphs, preventing orphaned headings that lose their explanatory context.

By respecting structural markers, each chunk becomes a self-contained semantic unit. A chunk might encapsulate an entire section on 'Deployment Prerequisites' including its introductory sentence, the subsequent bulleted list of requirements, and the closing note. This contrasts with fixed-length chunking, which could split the list arbitrarily, rendering the partial list meaningless to a retrieval system. The result is higher chunk coherence and reduced chunk contamination.

Structural chunking enables automatic metadata enrichment. A chunk derived from a section titled 'Security Protocols' can automatically inherit that title as a metadata attribute. This allows for filtered retrieval queries, such as searching only within 'Security' sections. The chunk's position in the document hierarchy (e.g., doc > section 3 > subsection 3.1) becomes a powerful, structured facet for hybrid retrieval systems that combine vector search with metadata filtering.

A common implementation is Markdown-Aware Splitting, where the parser uses Markdown syntax as explicit delimiters. The splitter is configured with a prioritized list of separators:

• ## (H2) for major sections
• ### (H3) for subsections
• \n\n for paragraph breaks
• | for table rows This ensures code blocks, tables, and nested lists remain intact as indivisible blocks, preserving their syntactic meaning for both the embedding model and the final LLM prompt.

The chunk size is controlled not by an arbitrary token count, but by the minimum header level specified. Setting the splitter to break on H2s creates large, comprehensive chunks covering entire top-level sections. Configuring it to split on H3s or H4s produces smaller, more granular chunks focused on specific sub-topics. This provides intuitive granularity control that aligns with the document's own logical organization, making it ideal for parent document retrieval strategies.

Structural chunking is highly dependent on the presence of explicit structural markers. It performs poorly on unstructured or poorly formatted documents that lack clear headings, such as raw text transcripts, OCR'd PDFs, or legacy HTML without semantic tags. In these cases, the algorithm may default to paragraph-level splitting or fail to create meaningful chunks, necessitating a fallback to semantic chunking or recursive chunking methods to impose logical boundaries where none exist.

A quality metric measuring whether a text segment contains a logically complete and self-contained idea without requiring external context to be understood. A coherent chunk can stand alone as a unit of information. Structural chunking aims to maximize coherence by using headings and sections as natural boundaries, avoiding splits that leave a chunk starting with 'On the other hand...' or ending mid-thought.