What the documents look like
Marketplace category access happens before listing or onboarding. The core material set is entity qualification and the authorization chain. A business license, trademark registration certificate, brand authorization chain and industry license form a group of documents that reference one another. Any difference in entity name, permitted scope or validity period across those documents affects the access decision. In cross-border material and multilingual retrieval, the same document may appear in several language versions, and a retrieval result needs to lead back to a checkable original text.
Material updates are irregular and are handled at each onboarding event. That frequency means a retrieval system cannot rely on a fixed synchronization rhythm. Each citation should be judged against the version submitted at that time.
Constraints at the citation tracing step
A knowledge base search node can associate one or more knowledge bases that share the same vector model, for vector search. Its output is an array of cited content, and the length may be zero. Even when no content is found, the output path still completes. The citation limit controls, in tokens, the maximum amount of content cited per search. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used; it is a filter threshold, not a sorting rule.
Retrieval results are handed to the answer stage, where the answer is checked line by line against the cited original text. Retrieval by itself does not constitute an authorization decision and does not confirm entity qualification. Citation tracing requires that every factual statement in an answer corresponds to wording in the cited original text returned by retrieval, and that the original text belongs to the currently valid document version.
Hypothetical exercise sample
Material original text for checking (fictional, used only for this page):
Authorizing party: Beichen Trading Co., Ltd. Authorized party: Xingye E-Commerce Co., Ltd. Authorized brand: NORTHFIELD. Authorization period: 1 January 2025 to 31 December 2025. Authorization scope: online sales in mainland China.
Draft sentence to be judged:
Xingye E-Commerce Co., Ltd. has obtained authorization from Beichen Trading Co., Ltd. for online sales of the NORTHFIELD brand in mainland China in 2026.
Check point: the authorization period in the original text ends on 31 December 2025, while the draft sentence asserts a 2026 authorization. The entities, brand and region match; the authorization period field differs. The draft sentence is unsupported by the material on the period field.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select only qualification material bases already confirmed to use the same vector model | The same vector model is the precondition for joint association; check knowledge base identifiers one by one and exclude those whose vector model differs |
| Search mode | Hybrid retrieval | The authorization chain contains numbers and proper nouns, and full-text retrieval can match exact fields; compare recalled content with a pure semantic retrieval result |
| Citation limit | Set to 3000 tokens for this exercise | This value is a hypothetical starting point on this page, used to observe citation trimming; compare the number and completeness of citations with a 1500-token result |
| Minimum relevance | Set to 0.4 for this exercise | This value is a hypothetical starting point on this page; compare which candidate items are filtered out against a 0.6 result |
| Question optimization | Turned off for this exercise | A single clear question needs no reference resolution; compare retrieval time and recalled content with the enabled setting |
| Result reranking | Turned on for this exercise | Secondary sorting when candidates are numerous; compare ordering with the disabled setting and observe changes in the first cited item |
Three common errors
- Treating cited content returned by retrieval as evidence that authorization is already established. Retrieval returns fragments of the material original text. Whether authorization holds requires checking the entity, object, period and scope fields in that original text. The retrieval action itself produces no authorization effect.
- Treating minimum relevance as a permission or entity filter. Minimum relevance is a 0-1 numeric filter threshold that removes low-relevance results only. It does not distinguish business entities and does not exclude unauthorized records. A high similarity score does not mean the material belongs to the current applicant.
- Writing a pending-review state as approved or already effective. When material is at the pending-review or pending-correction stage, it can only be recorded as awaiting verification. Signing, reaching cooperation and contract effectiveness are three distinct facts and cannot stand in for one another.
How to verify the setup
- Input the material original text and the draft sentence from the hypothetical sample above, and compare the authorization period field. The original text states 1 January 2025 to 31 December 2025, while the draft sentence asserts 2026. If the check conclusion is that the draft sentence is unsupported by the material on the period field, the check passes; if the conclusion is that the draft sentence is supported, it does not pass.
- Input the material original text from the hypothetical sample above and check whether the retrieval output returns cited content as an array. If the output is an array of length greater than 0 containing fragments traceable to the original text, citation output is normal; if the output is an empty array, record it as no recall for that instance and check whether the knowledge base has completed training. Do not conclude from this that the material does not exist.
- Input a configuration that associates only a single-vector-model knowledge base, and check whether any attempt is made to associate across different vector models. If all selected knowledge base identifiers match, the association condition is satisfied; if knowledge bases with differing vector models are found jointly associated, exclude the non-conforming ones first, record the actual observation, and do not run cross-model retrieval to compare missed recalls.