What the documents look like
Marketplace category access takes place before a listing goes live or a store is onboarded, and the review centres on the operating entity's qualifications and the chain of authorization behind the brand. The material types handled on this page are the business licence, the trademark registration certificate, the brand authorization chain and the industry licence. Together these determine whether a qualification statement inside a piece of product copy can be supported by the material already on file.
Updates to this material are irregular and each onboarding collects it again, so a retrieval result reflects the version of the material as stored. Before drafting, confirm whether the cited material covers the entity and the period named in the sentence being drafted. A brand authorization chain usually spans several documents, with the authorization tier, the authorized product scope and the validity period distributed across different passages. Retrieval can easily return only one tier of that chain, which leaves a drafted sentence citing an incomplete authorization relationship.
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 built on different vector models cannot be associated together; check the vector model identifier first, exclude those that do not match, and record what was actually observed. Semantic retrieval relies on vector distance, full-text retrieval suits precise queries such as numbers, models and proper nouns, and hybrid retrieval fuses the two paths through RRF. RRF is a fusion process, not a separate switch that can be turned on. The minimum relevance value is a filtering threshold that takes effect only under semantic retrieval or when result reranking is used, and it does not perform permission or entity isolation; it cannot be used to infer that unrelated businesses or unauthorized records were excluded. The citation limit is measured in tokens and controls the maximum amount of content cited per search. Image retrieval requires an image input and the corresponding model capability, and reranking requires a rerank model enabled through its settings; both are optional paths and do not run automatically on every request. Retrieval itself does not constitute authorization, and whether the material supports a given promotional sentence still requires manual comparison.
Hypothetical paper exercise sample: the source material reads, "Party A, a certain brand management company limited, authorizes Party B, a certain trading company limited, to use registered trademark No. 1234567 on Class 25 apparel goods, with an authorization period from 1 January 2024 to 31 December 2025." The sentence to be drafted reads, "All apparel sold in this store is officially authorized brand merchandise." The question to be decided is whether the phrase "officially authorized" in that drafted sentence can be supported by the source material above. The material states the authorizing party, the authorized party, the goods class, the trademark number and the authorization period. It does not state any basis for the word "official", and it does not state whether the trademark registrant is the authorizing party itself.
Configuration and implementation checks
The configuration items below relate to the sample above and list only documented settings or implementation checks relevant to this feature and the copy drafting task.
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Qualification material base | Check that the vector model identifier matches across bases before associating; exclude any base whose identifier differs |
| Search mode | Hybrid retrieval | The sample contains precise fields such as the trademark number alongside authorization relationship text; compare semantic, full-text and hybrid recall differences |
| Citation limit | To be determined, set provisionally by sample material length | Measured in tokens; compare authorization chain passages of different lengths to see whether content is truncated |
| Minimum relevance | To be determined, annotate relevant and irrelevant samples first | Takes effect only under semantic retrieval or result reranking; compare results before and after filtering against the annotation |
| Question optimization | Enable as needed | Compare the original question with the optimized question and check whether the authorization subject was altered |
| Result reranking | Enable as needed | Requires a rerank model enabled through its settings; compare scores and ordering before and after reranking |
The starting points above are settings for this hypothetical exercise and are not official default values. Numeric values must be determined through comparison against actual material.
Three common errors
- Treating a semantic recall result as a permission decision. Retrieval returning a piece of authorization material does not mean that material covers the subject or period of the current drafted sentence. In the sample the authorization period ends on 31 December 2025, so if the drafted sentence targets a 2026 sales scenario, a retrieval hit does not constitute support. Check method: input the drafted sentence and the source material, compare the authorization period with the period the draft applies to, and judge the draft unsupported where the periods do not overlap.
- Forcing knowledge bases built on different vector models into one association. The documentation states that associated knowledge bases must use the same vector model, and those with different identifiers cannot be associated together. Check method: read the vector model identifier of each base, exclude any base whose identifier differs and record it, and do not run a cross-model retrieval comparison.
- Treating RRF or the minimum relevance value as an isolation mechanism that can be switched on. RRF is the fusion process for hybrid retrieval results, and the minimum relevance value is a filtering threshold; neither excludes unauthorized records. Check method: input a query containing an unrelated business name, compare the returned results with the expected authorization scope, confirm that the filtering threshold did not perform entity isolation, and record the missing authorization determination as pending verification.
How to verify the setup
- Input the sample source material and the drafted sentence "All apparel sold in this store is officially authorized brand merchandise." Compare the authorizing party, the authorized party, the goods class, the trademark number and the period stated in the material, and observe whether the drafted sentence falls within the material's scope item by item. Where the material states no basis for the word "official", judge the sentence unsupported and record the gap as pending verification, without judging it as a denial or a conflict.
- Input a query containing only the trademark number, run it under semantic retrieval and under full-text retrieval, and compare the two sets of returned content. Observe whether full-text retrieval recalls the material passage containing that number, and whether semantic retrieval returns records with similar but different numbers. Use the difference between the two as the basis for choosing a search mode, without guaranteeing a first-hit result or complete recall.
- Input a clearly labelled paper empty-reference result and check whether the downstream copy drafting step handles it as an empty result. Observe whether the downstream step still generates a qualification statement when no citation is present, and record a downstream design gap if it does. This check does not claim that a fictional rule retrieval necessarily returns an empty array; it only uses a paper input to verify the downstream handling logic.