Marketplace promotions: citation tracing for product questions

Review citation tracing for product questions in Marketplace promotions. Check document scope, failure paths and source evidence through controlled tests.

What the documents look like

Marketplace promotion enrollment and price management turns on price, stock and campaign cadence. The work sits apart from product title compliance review, which examines promotional wording. This page covers citation sources and tracing inside product material and selection questions. Practitioners ask questions against four recurring records: the campaign enrollment form, the promotional price approval sheet, the stock lock sheet and the campaign retrospective sheet. Data updates weekly and once per campaign. A citation must lead back to a specific document version.

Two features of these records shape the tracing step. First, the approval sheet and the stock lock sheet carry explicit status fields, and those statuses move through stages that are not interchangeable. A record that has been retrieved is not a record that has been approved or executed. Second, enrollment forms, approval sheets and lock sheets are versioned by campaign, so the same product can appear with different values across versions. A citation that omits the version cannot be reconciled against a later review.

The knowledge base search node accepts one or more knowledge bases that share the same vector model. It outputs cited content as an array, and that array length may be zero. When no content is found, the output path still completes. Retrieval does not grant authorization. Whether an enrollment form, approval sheet or stock lock sheet stands at a given stage must be confirmed against the source document.

Constraints at the citation tracing step

The search node output is an array of citations. Downstream, the answer stage consumes that array. Each answer sentence that relies on a citation must be checked against the cited original text for subject, object and status.

A hypothetical drill sample follows. On 10 June 2025, a marketplace campaign operator submits a promotional price approval sheet. The original text reads: "Campaign name: 618 promotion; applicable product: model X air fryer; proposed enrollment price: 299 yuan; approval status is pending review; submitter: operator A." The stock lock sheet from the same batch reads: "Locked quantity: 500 units; lock status is pending confirmation." A handover request draft sentence reads: "This product has completed enrollment at 299 yuan, and 500 units are locked, ready to list."

The question to decide is whether that draft sentence is supported by the materials. It is not. The approval sheet states pending review, and the lock sheet states pending confirmation. The draft sentence asserts completed enrollment and a completed lock. A pending request is not approval, and an offered option is not execution. The materials support neither claim.

Relevance scores and semantic retrieval do not act as permission or entity filters. A similarity setting does not exclude unrelated businesses or unauthorized records. The same-vector-model requirement for jointly associated knowledge bases does not extend to business registration sites or other external systems.

Configuration and implementation checks

Configuration or check itemDocumented contentImplementation check on this page
Associated knowledge basesOne or more knowledge bases sharing the same vector model may be selectedCompare the vector model identifier of each selected base; exclude any that differ, and do not run a cross-model search
Search parametersCitation cap is controlled in tokens; minimum relevance applies only under semantic retrieval or result rerankingRecord the assumed trial starting value and the value it is compared against
Citation outputArray format, length may be zeroUse a paper input of an empty array to check whether the downstream path still completes
Result rerankingReRank performs a second ordering of candidates, producing a 0-1 scoreConfirm a rerank model is present and enabled per settings; where it is not enabled, record the actual observation
Citation handoffSearch output passes to the answer stageCompare each answer sentence against the cited original text for subject, object and status
Material timelinessTesting sources against current and historical versionsRecord the cited version; whether it is adopted still depends on review against the effective version

The assumed trial starting value for the citation cap is a paper figure chosen for this drill only, compared against a lower cap to observe whether the cited span shrinks. It is not an official default. The thinking budget is a reasoning budget and is not used as a citation token cap.

Three common errors

  1. Treating a retrieval hit as approval. Input: the approval sheet above, whose approval status field reads pending review, is retrieved. Compare: the retrieval result against the approval status field in the source text. Observable result: if the answer states that enrollment is complete, that conflicts with the source field value pending review, and the statement is unsupported.
  2. Treating a pending lock as a completed lock. Input: the stock lock sheet original text reading lock status is pending confirmation. Compare: the draft sentence stating the lock is complete against the source status field. Observable result: the two do not match, and the draft sentence is unsupported.
  3. Forcing association of knowledge bases with different vector models. Input: two knowledge bases whose vector model identifiers differ. Compare: the model identifiers. Observable result: where the identifiers differ, the bases are excluded from joint association, and no cross-model search is run to compare results.

How to verify the setup

  1. Citation handoff check. Input: a paper input of an explicitly empty array. Compare: whether the downstream answer stage still completes. Observable result: the path completes without error, which matches the documented behavior of a zero-length citation array. If the path breaks, record it as an open item.
  2. Citation-to-source consistency check. Input: the approval sheet original text above and the draft sentence. Compare: subject, object and status across the three fields. Observable result: the draft sentence claims completed enrollment while the source reads pending review, so the sentence is judged unsupported. A pending request is not approval.
  3. Material version check. Input: the current version and a historical version of the same document. Compare: the value of the same field in both versions. Observable result: where the values differ, record the cited version and route it to effective-version review; where the values match, that match does not by itself indicate authorization.

Across all three checks, the same subject, the same period and the same field must be compared. Only explicitly mutually exclusive values in one field constitute a conflict. Where a field is absent, or where only evidence of a related stage exists, the result is unsupported or open, not a denial or a conflict. Missing evidence remains missing.