Zhihe GrowthA request for indoor, 220 V AX-220 documentation is not equivalent to “recommend similar cleaning equipment.” The broader wording may return more material while dropping the model and electrical constraints. Counting that as improvement corrupts both procurement accuracy and GEO test comparability.
Extract protected fields without discarding the original
Do not silently continue with broadened constraints after validation fails. A rewrite that drops EU scope, 220 V or an outdoor exclusion may retrieve more records while making the recommendation invalid. Retain the original query and changed fields for review. If relaxation is appropriate, state it explicitly and separate relaxed candidates from records meeting the original requirements. Similar model names, higher similarity scores or more results do not authorize overriding constraints.
Keep the complete original question and an immutable question ID. Identify entity, region, values and units, mandatory conditions, exclusions and requested output. Extraction itself can be wrong: the structured record is a reviewable interpretation, not a replacement for the original. Language does not establish the buyer's region.
| Field | Example | Permitted change | Change to reject |
|---|---|---|---|
| model | AX-220 | Approved alias with identity retained | AX-110 |
| region | EU | Equivalent explicit wording | Worldwide |
| voltage | 220 V | Standard spacing | 110 V or omission |
| excluded | outdoor | Preserve the prohibition | Recommend outdoor use |
| task | Verify supply conditions | Equivalent phrasing | Recommend a brand instead |
Inspect three executed comparisons
The constraint_diff function in controls.py compares protected fields. One example retains all fields, one changes the model to AX-110 and one removes the exclusion. The query_constraints output in results.json identifies model and excluded respectively.
{
"model": "AX-220",
"region": "EU",
"voltage": "220 V",
"excluded": ["outdoor"]
}
This deterministic gate does not infer semantics from arbitrary text or run a rewriting model. It demonstrates what to check after an editor or model produces a candidate. If extraction reverses a prohibition in both records, field equality will not discover that error. Review against the original remains necessary.
Synonym expansion is not compatibility expansion
Approved Chinese names, English abbreviations and standard identifiers can support alternate queries. Nearby model families cannot automatically become acceptable substitutes. Hyphens, unit prefixes and negative conditions deserve particular attention.
The Stanford information retrieval text supplies retrieval foundations, not domain-specific compatibility rules. Maintain an explicit terminology table, evidence for aliases and prohibited substitutions. Replacing “power-supply input” with “motor voltage” changes the attribute rather than improving its wording.
Make fallback behavior explicit
| Finding | Next action | Retained record |
|---|---|---|
| Protected fields agree | Retrieve alongside original | Variant and generator version |
| Unsupported new constraint | Reject and use original | Added field and reason |
| Exclusion disappears | Block automated answer | Original negative phrase and diff |
| Original is ambiguous | Clarify or branch explicitly | Unresolved fields |
| User approves broader scope | Assign a new question ID | Approval and relationship |
Do not delete the original because a rewrite retrieves more candidates. The extra candidates may be irrelevant. Safety, compatibility and certification questions require evidence of constraints, not a maximum result count.
Evaluate benefit without changing the test
Freeze the corpus and relevance key, then compare original queries with accepted variants using recall, ranking and constraint retention. Preserve rejected variants and aggregate paired results by question. A changed requirement needs a changed relevance key; it is not a fair comparison against the old one.
For commercial AI-platform retesting, preserve the original wording, session conditions and mode. Experimental rewrites belong in a separate analysis. Selecting only formulations that elicit the brand and adding them to the original citation rate is outcome selection. Internal retrieval quality and public website citation are different conclusions.
Zhihe Growth's application and acceptance
When organizing Chinese exporters' bilingual question sets, Zhihe Growth should make product identifiers, conditions and procurement constraints auditable field by field. Acceptance can inspect the original question, candidate variants, rejection reasons and selected version, with additional attention to numbers and negation.
Continue with intent labels, bilingual terminology and cross-language retrieval. The technical FAQ supplies direct answers; the advanced package includes the executable comparisons and review forms. Neither is a guarantee of unmeasured citation growth.