What the documents look like
Marketplace category qualification happens before a listing is published or a store is onboarded. The core material is the operating entity's qualification and the chain of authorization behind the brand. For stock procurement and supplier files, the documents that need to be retrieved are the business license, the trademark registration certificate, the brand authorization chain, and industry licences.
A recurring trait of these materials is that the same document type appears across different suppliers, different categories and different validity periods, and an authorization chain frequently crosses several layers of entities. The retrieval step has to answer whether a given document can support a given category, a given period and a given entity. Materials are resubmitted on an irregular basis, once per onboarding or once per added category. The retrieval library has no automatic synchronization relationship with external public registry systems. A knowledge base search node can be associated with one or more knowledge bases that use the same vector model; this is a precondition for association.
Constraints at the retrieval step
The retrieval difficulty for authorization chain material is that the entity, the object and the period have to align at the same time. A retrieval result that matches a passage of text only shows that the passage is close to the query in vector distance or keyword terms. It does not show that the material is valid under the current authorization state.
Semantic retrieval relies on vector similarity and suits natural language descriptions. Full-text retrieval relies on keyword hits and suits numbers, models and proper nouns. Hybrid retrieval uses both recall paths and merges them through RRF. Result reranking uses a ReRank model to reorder candidate text and produces a score between 0 and 1. The minimum relevance value is a filter threshold and takes effect only in semantic retrieval or when result reranking is used; it does not perform permission filtering or entity isolation. The reference cap is measured in tokens and controls how much content a single search may reference. Question optimization performs coreference resolution and question expansion and adds one model call before the formal retrieval.
Hypothetical paper exercise
Material text for comparison (fictional, used only for this page):
"Party A: Xinghai Supply Chain Management Co., Ltd. Party B: Yunling Food Co., Ltd. Party A authorizes Party B to sell 'Yunling' brand pre-packaged food in the East China region between 1 January 2025 and 31 December 2025. This authorization letter does not cover any trademark other than trademark registration number 0000000."
Draft sentence to be assessed: "Xinghai Supply Chain Management Co., Ltd. has authorized Yunling Food Co., Ltd. to sell all Yunling brand goods nationwide on a long-term basis."
The source states the region as East China, the period as 1 January 2025 to 31 December 2025, the object as pre-packaged food, and the trademark as limited to number 0000000. The draft sentence changes the region to nationwide, the period to long-term, and the object to all goods. All three points exceed the scope stated in the source.
Configuration and implementation checks
| Configuration item | Starting point for this hypothetical exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the qualification library that uses the same vector model | The official description requires the same vector model; check each library identifier one by one and exclude those whose model does not match |
| Search mode | Run the sample once with semantic retrieval and once with full-text retrieval | Compare how the two modes recall fields such as authorization region and authorization period, then decide whether to switch to hybrid retrieval |
| Reference cap | Start from a value large enough to hold one complete authorization letter body | The unit is tokens; test with authorization letters of different lengths to see whether content is truncated, and adjust accordingly |
| Minimum relevance | Label relevant and irrelevant samples first, then set the value from those labels | The value takes effect only in semantic retrieval or when result reranking is used; compare whether irrelevant passages are filtered out at different values |
| Question optimization | Enable for follow-up questions that contain references, disable for queries with explicit fields | Compare the original question with the optimized question and check whether the entity or the object was altered |
| Result reranking | Enable when candidates are numerous and check the rerank score | The rerank score runs from 0 to 1; compare the ranking position of the same candidate before and after reranking |
Three common errors
- Treating a retrieval hit as a completed authorization. In the sample, the draft sentence states nationwide, long-term sales of all goods. Retrieval can recall the authorization letter, but the source covers only East China, the single year 2025, and pre-packaged food. The input is the draft sentence and the authorization letter source, the comparison covers the region, period and object fields, and the observable result is that all three points exceed the source scope, so the statement is unsupported.
- Using minimum relevance to exclude unrelated entities. Minimum relevance is a similarity filter threshold and does not perform entity isolation. In the sample, if an authorization letter from another company is retrieved, it can be recalled even at a high similarity. The input is the query and the candidate list, the comparison covers the entity names in the candidates, and the observable result is the appearance of an entity that is neither Xinghai nor Yunling, which requires manual exclusion rather than reliance on the threshold.
- Treating a pending correction as approved. Suppose a supplier submits a new industry licence correction and its status is pending review. The input is the correction and the review record, the comparison covers the status field, and the observable result is that the status remains pending, so the supplier's category qualification cannot be confirmed on that basis.
How to verify the setup
- Run one retrieval comparison using the sample draft sentence and the authorization letter source. The input is the two texts above, the comparison covers the values of the region, period and object fields, and the observable result is that all three points in the draft sentence exceed the scope stated in the source. If the retrieval result directly supports the draft sentence wording, the reference content from that retrieval has been misinterpreted and the source needs to be rechecked.
- Test the recall scope with a query whose entity does not match. The input is "trademark registration certificate of Yunling Food Co., Ltd.", the comparison covers the entity names in the returned candidates, and the observable result is that certificates of the same type belonging to other entities may appear. If results include a non-Yunling entity and it carries no manual annotation, record the observation and check whether the associated knowledge base identifiers are consistent.
- Check downstream handling with an empty-result input. The input is an explicit empty reference array, the comparison covers whether the downstream node continues along that output path as described in the documentation, and the observable result is that the path still executes when the reference content length is 0. This check covers only the handling of an empty result and makes no assertion about whether a fictional rule returns an empty set.