Chunk Boundary

Retrieval operates on fragments, not documents, so where a document is cut decides what a model can see. A boundary placed mid-argument produces a passage whose meaning depends on text that was not retrieved; a boundary that ignores structure can separate a rule from its exception, which is the failure mode with real consequences in policy and regulatory content. Fixed-length chunking is simple and structure-blind. Structure-aware chunking respects headings, clauses and record boundaries, producing fragments that remain true when read alone. The test is whether a retrieved passage would mislead a careful reader who saw nothing else.

Centralpoint indexes at the record level, so the unit of retrieval corresponds to a governed object with its own classification, audience assignment and version history rather than to an arbitrary span of characters. A retrieved fragment therefore arrives with the same access rules and provenance as the record it came from.