On this page

The short answer

Test whether a child chunk finds the right evidence and whether parent expansion preserves its meaning without introducing ineligible or distracting text. Keep both stages visible in the evaluation record.

What to take away

  • Preserve child-to-parent lineage.
  • Recheck eligibility after expansion.
  • Measure the added context budget.

Sources checked . Research synthesis and Rubrex recommendations. Examples and local experiments are labeled; they are not client results.

Define the expansion contract

Azure describes document-to-chunk indexing patterns and structure-preserving chunking. In a parent-child design, a smaller unit may determine the match while a larger unit supplies generation context. Specify the relationship explicitly: document ID, revision, parent section, child offsets, and permission metadata. A successful child match is not sufficient if the parent lookup returns an obsolete version or a broader document with different access rules. Evaluate expansion as its own transformation.

Evidence: Chunk large documents for RAG and vector search [1]Document Layout skill [2]

Build a parent-version trap

Author a handbook section whose child paragraph states a delivery deadline and whose parent heading limits the rule to domestic orders. Create a revised parent with a different international exception. Ask one domestic and one international question. A child-only result may omit the scope; the correct parent supplies it. A mismatched parent revision produces a plausible but incorrect rule. The acceptance criterion requires the child and parent to belong to the same intended source revision and eligibility scope.

Measure what expansion contributes

For each retrieved child, save its rank, supporting span, parent ID, parent revision, and expanded context length. Score whether expansion restores required qualifiers and whether it displaces other necessary sources under a fixed total budget. Count parents returned repeatedly from several child hits only once when measuring distinct evidence. Inspect whether a single large parent dominates the context even though another source is required. Separate a child retrieval miss from a failed parent lookup or an overly broad expansion.

Choose an evidence-bounded expansion policy

Prefer the narrowest parent that restores the question’s required context. Test section-level expansion against full-document expansion rather than assuming more context is always useful. Reject expansion when lineage or eligibility cannot be established, and expose that failure to the application instead of silently substituting a nearby document. Keep a control question whose child already contains the complete rule; it reveals whether expansion adds unnecessary cost or distractors without improving answer acceptance.

Measure expansion failures independently from answer failures so missing parents remain visible.

Account for shared parents before budgeting context

Several matching children can point to the same parent section. Resolve those links before assigning context slots so repeated parent expansion does not crowd out another required document. Retain which child hits justified including the parent even when its text appears only once. That lineage helps explain why a large section was selected and whether the matched evidence was actually present. Add a fixture with three child hits from one parent and one hit from an independent exception document; the acceptance rule still requires the exception to survive final context assembly.

Limits of the evidence

This is an architecture test method, not a claim that a particular framework implements all lineage or permission checks automatically. Examples are synthetic.

Common questions

Should the parent always be the whole document?

No. Compare the smallest expansion that restores the missing context with the larger alternative.

Why keep revision IDs on children?

They let the evaluator detect a correct child joined to a stale or incompatible parent.

Sources & further reading

Primary sources behind this briefing. A source’s findings apply to its own study conditions; publication on arXiv does not establish peer review.

  1. Chunk large documents for RAG and vector search Microsoft · Living reference · Maintained documentationReviewed: chunking strategies and overlap; no copied universal optimum. Accessed October 16, 2026.
  2. Document Layout skill Microsoft · Living reference · Maintained documentationReviewed: structural extraction, supported representations and location metadata. Accessed October 16, 2026.

Questions or corrections? Write to Rubrex. Read our editorial approach.

From reading to a repeatable evaluation

Put a quality bar around your use case.

The AI Reliability Sprint covers one use case, a baseline, failure analysis, and one improvement experiment in 10 business days.

Request an evaluation call

See the Sprint deliverables →