What the documents look like
Category access review happens before a product is listed or a store is onboarded. The core materials are the business license, trademark registration certificate, brand authorization chain, and industry license. These are collected by marketplace development staff, store onboarding specialists, or brand channel managers, and they come from marketplace recruitment pages, the National Enterprise Credit Information Publicity System, trademark office data, and license systems. Updates are irregular, and materials are resubmitted at each onboarding.
For knowledge base retrieval and recall, these materials share a common trait: subject names, certificate numbers, validity periods, authorization tiers, and business scope fields are dense. Questions typically ask whether a subject holds a certificate or whether an authorization chain covers a category. Retrieval results need to locate the specific source material for manual review. Retrieval itself does not decide whether a qualification is valid. The update frequency does not change the retrieval step, so it is not treated as a configuration factor 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 with different vector models cannot be associated together. Search operates at the whole-library level; collections are only for classification and do not affect search results. A single retrieval combines problem optimization, semantic search, full-text search or hybrid search, image retrieval, and reranking according to the input and search parameters, then produces reference content after similarity filtering and reference limit trimming. Reference content is output as an array and can have a length of 0, meaning the output path still runs when nothing is found.
Semantic search relies on vector distance, full-text search relies on keyword hits, and hybrid search merges the two paths through RRF. RRF is a result fusion process, not a switch that can be turned on separately. The minimum relevance score takes effect only in semantic search or when result reranking is used. It is a filtering threshold and does not perform permission or entity isolation. Recalling a material does not mean the qualification it describes has been approved or the authorization is in force. The same-vector-model requirement applies only to knowledge base association; it does not extend to external systems such as business registration or trademark databases.
Hypothetical paper exercise sample: One material in the knowledge base reads: "Authorizing party: Jia Brand Management Co., Ltd.; Authorized party: Yi Trading Co., Ltd.; Authorized brand: Xingye; Authorized category: household appliances; Authorization period: January 1, 2025 to December 31, 2025." Another business license reads: "Name: Yi Trading Co., Ltd.; Unified Social Credit Code: 91XXXXXXXXXXXXXX; Business scope: household appliance sales." A draft sentence to be checked reads: "Yi Trading Co., Ltd. has obtained authorization for the Xingye brand household appliance category, valid until December 31, 2025." The check compares the draft sentence's subject, brand, category, and period against the authorization material. The business license only supports the subject's existence and business scope; it does not prove the authorization relationship.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the same-vector-model library holding the source qualification materials | First read the vector model identifier of each library, exclude those with different identifiers, then decide the association scope |
| Search mode | Start with hybrid search | Materials contain certificate numbers and natural language descriptions; compare recalled content from semantic, full-text, and hybrid modes on the same test question |
| Reference limit | Set to a token amount that can hold one complete authorization chain | Test with actual materials of different lengths to see whether reference content is truncated, then adjust the limit |
| Minimum relevance score | Start with a middle value, then adjust after labeling relevant and irrelevant samples | This value takes effect only in semantic search or result reranking; compare whether irrelevant fragments are filtered at different values |
| Problem optimization | Enable for multi-turn follow-up scenarios, disable for single clear questions | Compare the original question with the optimized question, and check whether subject, brand, or category was altered |
| Result reranking | Enable when a rerank model is available and candidates are numerous | Rerank scores can be used for filtering; compare ranking and minimum relevance interaction before and after reranking |
Three common errors
- Treating a retrieval hit as proof that authorization exists. Input "Yi Trading Co., Ltd. Xingye authorization"; the retrieval returns the authorization material. This only shows the material was recalled, not that the authorization was signed or is in force. The check compares the draft sentence against the material's subject, brand, category, and period fields for same-period, same-field consistency. Signing status not stated in the material should be recorded as pending verification.
- Associating knowledge bases across different vector models. Associating two libraries that use different vector models violates the node's same-vector-model requirement. The check reads each library's vector model identifier, excludes those with different identifiers, and records the actual observation. It does not infer runtime results of the association.
- Using the minimum relevance score as a substitute for permission filtering. The minimum relevance score is a similarity filtering threshold and does not distinguish business entities or authorization scope. The check confirms that the retrieval scope is already limited to libraries with access rights. When access rights are missing, record the evidence gap.
How to verify the setup
- Input the draft sentence and the two source materials from the hypothetical sample. Compare the draft sentence's authorization period field against the authorization material's "January 1, 2025 to December 31, 2025." If the draft says "valid until December 31, 2025," the period endpoint matches the source, but the source does not state signing or effectiveness status, so that part remains unsupported. If the draft says "valid long-term," it contradicts the explicit period in the source, and the check records an inconsistency.
- Input a short question containing only the subject name and observe the reference content array. If the returned length is greater than 0 and the content points to materials related to that subject, recall is usable. If an empty array is returned, treat the empty result path as documented and use it as a paper input for downstream design review. Do not claim that a particular rule will necessarily return an empty result.
- Input an exact query containing a certificate number and compare recalled content from full-text search and semantic search. If the number query hits the source material under full-text search, that mode fits exact queries. If the two modes return different content, record the difference and check whether problem optimization altered the number or subject. Do not guarantee a first-position hit or complete recall.