Product listing reviews: retrieval for operating reviews

Review retrieval for operating reviews in Product listing reviews. Check document scope, failure paths and source evidence through controlled tests.

What the documents look like

Title and detail-page compliance work turns on four document families: the product title table, detail-page assets, the prohibited-advertising-terms list, and quality inspection reports. Together these decide whether a proposed claim can go live. The judgment rests on advertising truthfulness, qualification display, and screening for prohibited wording. Parcel delivery is not part of this work.

Inspection reports typically name the testing body, sample name, test items, and report number. Their validity period and the models they cover define the boundary of what they support. The prohibited-terms list states banned or restricted expressions. The title table and detail-page assets are the objects being checked. Retrieval answers a narrower question: whether a proposed sentence is related to a list entry, a report conclusion, or a platform rule statement, so a human can review it.

Materials update weekly, and each new product launch produces new versions. Operating reports and anomaly reviews therefore need to separate title and asset versions for the same product across different weeks and launch batches. Retrieval does not resolve which version governs; that remains a human determination.

Constraints at the retrieval step

The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases with different vector models cannot be associated together. The node outputs cited content as an array, and that array can have length 0. When nothing is retrieved, the output path still completes.

Semantic retrieval relies on vector distance. Full-text retrieval relies on keyword hits. Hybrid retrieval merges the two paths through RRF. ReRank performs a second ordering of candidate text and returns a 0–1 score. The minimum relevance setting takes effect only under semantic retrieval or when result reordering is used. The citation limit is measured in tokens. Question optimization performs coreference resolution and question expansion, and it adds one model call before retrieval.

These capabilities affect recall and ordering only. They do not verify a product, a qualification, or an authorization boundary. A retrieved prohibited-terms entry shows that the entry is related to the proposed sentence; a human still compares the entry's scope of application against the sentence's subject. A pending handover request is not an approval, and an offered option is not an executed action.

Configuration and implementation checks

The following hypothetical paper exercise illustrates a checking method. It refers to no real entity or record.

Material text: "The March 2024 edition of the Prohibited Advertising Terms List, Article 12, states that 'national-level' is a prohibited expression. Quality inspection report number JC-2024-0311 covers a rice cooker of a stated model, tests electrical safety, is dated 5 March 2024, and does not state any conclusion about advertising wording." Proposed draft sentence: "This product is a national-level certified rice cooker." Handover request: "Please confirm whether this title can be used, and state the basis."

Check itemHypothetical starting value for this exerciseReason for the choice and comparison object
Associated knowledge basesOnly the two bases holding the prohibited-terms list and the inspection reportsBoth use the same vector model, which satisfies the joint-association condition
Search modeRun full-text retrieval and semantic retrieval separately, once eachList entries are exact expressions; report conclusions are semantic content; observe the difference in hits
Citation limitSet to 800 tokens for this exerciseCompare against the actual length of the list entry and the report conclusion paragraph, then adjust
Minimum relevanceSet to 0.5 for this exerciseCompare against the score distribution of manually labelled relevant and irrelevant samples, then adjust
Question optimizationOff for this exerciseThe draft sentence has a complete subject; first observe recall on the original question
ReRankNot enabled for this exerciseThe candidate count is small; first observe ordering without reordering

Three common errors

  1. Treating a retrieval hit as a compliance conclusion. A returned prohibited-terms entry only shows that the entry is related to the draft sentence. Whether the wording violates the rule still requires a human comparison of the entry's scope against the sentence's subject.
  2. Forcing joint association of bases with different vector models. When two bases carry different vector model identifiers, they should not be associated together. Record the identifier mismatch and exclude the mismatched base. Do not execute a cross-model retrieval to compare recall, and do not infer missed recall, redundant returns, or unpredictable behaviour.
  3. Treating a pending review state as granted authorization. The handover request is pending confirmation. That state does not mean the title has been approved for use, and it does not mean operating permission has been obtained.

How to verify the setup

  1. Enter the material text and the draft sentence above, and run a full-text retrieval for "national-level" in the associated bases. Compare whether the returned citations include the original text of Article 12 of the prohibited-terms list. The check condition is met when that original entry appears. When it does not appear, record a miss. Do not infer that the list lacks the entry.
  2. Enter the inspection report number "JC-2024-0311" and run a retrieval. Compare whether the returned citations point to that report record, and check whether the report states a conclusion about advertising wording. When the report is returned but states no such conclusion, record the claim as unsupported. When an empty array is returned, carry the empty result into the downstream design for empty results. Do not claim that a rule retrieval necessarily returns an empty array.
  3. Enter the handover request and compare the retrieval result against the request fields. Confirm that the request is in a pending-review state. When no authorization or effective basis is attached, record it as pending verification. Do not judge it as approved or rejected.

Across all three checks, the judgment must concern the same subject, the same period, and the same field. Only mutually exclusive values in the same field constitute a contradiction. Fields that are not stated, or evidence that covers only a related stage, are recorded as unsupported or pending verification. A pending review does not demonstrate a lack of operating permission. Signing, reaching cooperation, and contract effectiveness are three distinct facts. A quotation validity period does not indicate a supply period. Do not supply authorization or effectiveness clauses to obtain a definite answer; keep the evidence gap explicit. Materials needed for a check must sit within an accessible target retrieval scope or in a clearly stated manual input. When they are missing, record the gap.