Changing document chunk size improves retrieval scores, but citations now point to large sections that do not clearly justify the answer. The search result changed its evidence boundary along with its size.
For teams tuning knowledge retrieval, inspect citation support after chunking changes. A better-looking aggregate score should not conceal lost qualifications or unusable source locations.
What does chunk size change in citation review?
It changes the text units available for retrieval and context assembly. Larger units may preserve surrounding conditions while adding irrelevant material; smaller units may improve focus while separating a rule from its exception.
The right boundary depends on document structure and questions. A table row, section heading and footnote relationship may need treatment beyond a fixed character count.
Compare the same source questions
Use fictional documents with reviewed expected passages. The proposed matrix checks support and provenance rather than only search ranking.
| Fixture | Citation property to inspect |
|---|---|
| Rule with nearby exception | Both relevant parts remain available |
| Multi-section policy | Cited location identifies the supporting section |
| Table-dependent answer | Header and row meaning remain associated |
| Repeated heading | Source identity is not confused |
| Unsupported detail near relevant text | Citation does not justify an invented claim |
Keep the document version, parser output and questions stable when chunking is the variable being evaluated. Otherwise a simultaneous extraction change can confound the result.
Review retrieval and citation precision separately
A retrieved chunk can contain the answer somewhere while the cited span points elsewhere. Conversely, an exact citation cannot help if the required passage was never retrieved.
The exact-claim citation guide defines useful review categories. The header and footnote guide identifies structural relationships that fixed-size splitting can disturb.
Preserve source coordinates through reindexing
Derived chunk identifiers may change after a new configuration is applied. Keep a stable relationship to the original document version and location so evaluation results and user citations remain interpretable.
My knowledge platform contribution supplies related project context, not a measured claim about an optimal chunk size. Any new comparison needs its own workload, configuration and limitations.
Choose the configuration from supported task evidence and operating costs. A larger context can increase processing work, while a smaller one can require more assembly. Neither trade-off should be hidden behind a universal recommendation that one chunk length is best.
Does higher retrieval recall guarantee better citations?
No. Inspect whether the cited evidence supports the exact claim and provides a usable location.
Should every document use the same chunk boundary?
Consider the relevant structure and supported tasks; fixed size alone may not preserve tables or qualifications.
What should we compare first?
Use a rule-and-exception fixture under both configurations in the retrieval evaluation discussion.