Marketplace promotions: retrieval for multilingual work

Review retrieval for multilingual work in Marketplace promotions. Check document scope, failure paths and source evidence through controlled tests.

What the documents look like

Promotion registration and price management in this category turns on price, stock and campaign rhythm. The working materials are campaign registration forms, promotional price approval sheets, stock lock sheets and campaign review sheets. These materials commonly exist in parallel Chinese and target-market language versions, and one campaign may carry registration conditions and price terms in more than one language at the same time. A retrieval task therefore has to match across languages, which places semantic retrieval's documented cross-multilingual understanding directly in scope.

Materials update weekly and per campaign. Historical transaction prices sit alongside current registration information, and the retrieval scope needs to cover data that has completed training. The documents do not state a batch synchronization, full synchronization, real-time or transaction-triggered mechanism, so none is assumed here.

Constraints at the retrieval step

The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases on different vector models cannot be associated together. During operation, verify the vector model identifier first, exclude those that do not match, and record the actual observation. Do not run a cross-model retrieval to compare missed recalls.

Cited content is output as an array, and its length can be 0, meaning the output path still runs even when nothing is retrieved. The minimum relevance setting takes effect only under semantic retrieval or when result reranking is used. It is a filter threshold and does not perform permission or entity isolation; retrieval results do not constitute authorization. Question optimization performs coreference resolution and question expansion and adds one model call.

Hypothetical paper exercise sample. A brand sales planner holds an English campaign registration form whose source text reads: "Campaign: Spring Home Sale. Registration deadline: 2025-03-10. Submitted price: 199 CNY per unit. Stock locked: 800 units." A separate Chinese promotional price approval sheet states: "Minimum final price during the campaign: 189 CNY. Approval status is pending review." The draft sentence to be judged is: "The campaign price has been approved and 800 units of stock are locked." The handover request to be checked is: "Please confirm whether this campaign can proceed to shipment."

Configuration and implementation checks

Configuration itemRelevant setting or check on this pageCorrespondence with the sample
Associated knowledge basesVerify each base's vector model identifier; associate only those on the same vector modelThe Chinese approval sheet base and the English registration form base must share a model
Search modeCompare semantic, full-text and hybrid retrieval against the test questionSemantic retrieval for the English form; full-text retrieval for approval status
Citation capUnit is tokens; test content completeness with materials of different lengthsHypothetical exercise starting value of 2000 tokens, chosen because the sample contains two language passages; compare completeness against 1000 tokens
Minimum relevanceTakes effect only under semantic retrieval or result reranking; label relevant and irrelevant samples firstLabel the "pending review" and "approved" samples separately
Question optimizationCompare the original question with the optimized question; check whether the subject was alteredCheck whether "this campaign" is completed as the correct campaign name
Result rerankingSecond-pass ordering when a model is available and enabled by configurationEnable when candidates are numerous; check score and subject consistency

Three common errors

  1. Treating "Submitted price: 199 CNY" on the English registration form as an approved price while ignoring the "pending review" status on the Chinese approval sheet. Check: input both source texts, compare the value of the same field for the same campaign, and observe whether an "approved" statement appears.
  2. Treating minimum relevance as a permission filter, so that a low-scoring result being excluded is read as meaning no permission to view it. Check: input a low-relevance record that falls inside the target scope, observe whether it is filtered by the threshold, and confirm that filtering and authorization are separate matters.
  3. Reading an empty citation array as a retrieval failure or as a business action being unavailable. Check: input a paper query with no match, verify whether the downstream design accepts an empty result, and draw no business conclusion from it.

How to verify the setup

  1. Input the two source texts and the draft sentence from the hypothetical sample, compare whether "pending review" and "approved" are conflated, and observe whether the draft sentence is marked as unsupported.
  2. Input the handover request "Please confirm whether this campaign can proceed to shipment," compare the retrieval result against the authorization boundary, and observe whether a retrieval hit is misread as executable authorization.
  3. Input a short query containing only the campaign name, compare the original question with the retrieval terms after question optimization, and observe whether the subject is wrongly expanded or replaced with a different campaign.

The approval status remains pending review, so the draft sentence's claim of approval is unsupported by the supplied materials. Whether shipment may proceed is not established by retrieval; the retrieval step returns reference content and does not grant authorization. Where a field is not provided in the documents, its value remains unknown.