What the documents look like
Category access work happens before a product is listed or a store is onboarded. The objects under review are the operating entity's qualifications and the chain of authorization behind the brand. In the inventory procurement and supplier materials task, the recurring documents are the business license, the trademark registration certificate, the brand authorization chain, and the industry license. These are collected by marketplace development staff, store onboarding specialists, or brand channel managers, and they enter a knowledge base for retrieval and citation. Materials are updated irregularly, and each onboarding attempt can produce a new version.
For vector models and indexes, the practical difficulty is that one supplier's license, trademark certificate, authorization letter, and permit are four separate documents. Entity name, registration number, licensed scope, and validity period sit in different fields, each with its own wording. A retrieval query can pull fragments belonging to a different supplier or a different legal entity into the same cited passage. The category difference from store review maintenance is that access control is decided before listing or onboarding, and the core question is the entity qualification and the authorization chain.
Constraints at the vector indexes step
Knowledge bases joined in one retrieval must use the same vector model, so the actual model and input type are checked first. One source data block can be associated with multiple indexes, and the index content is checked separately from the source content. Index content should be complete and focused; retrieval coverage and index length are weighed against sample material. Image retrieval also depends on whether image descriptions or image vectors were generated at ingestion time, and query-side behavior alone does not settle that.
A similarity score is a ranking basis. It does not perform entity isolation or permission filtering. The same-vector-model requirement applies to knowledge base association and does not extend to business registration sites or other external systems. RRF is the fusion step for mixed retrieval results, not a switch that can be turned on. The citation cap is measured in tokens, and the minimum relevance value takes effect only in semantic retrieval or when result reranking is used. The dataset search node outputs cited content as an array whose length can be zero, so a downstream path still runs when nothing is retrieved.
Assumed paper exercise sample: a supplier package contains a business license reading "Unified Social Credit Code 91310000MA0000000X, entity name: Example Supply Chain Management Co., Ltd., operating status: active, business scope includes sale of daily necessities." The brand authorization chain reads "Authorizer: Example Brand Management Co., Ltd.; authorized party: Example Supply Chain Management Co., Ltd.; authorized category: daily necessities; authorization valid until 31 March 2027; this authorization letter may not be sublicensed without the authorizer's written consent." The draft sentence to be judged is "Example Supply Chain Management Co., Ltd. has obtained complete authorization for the daily necessities category and may be submitted directly for platform onboarding." That draft treats the existence of an authorization chain as onboarding submission eligibility. The authorization letter does not state any platform onboarding submission right, so the conclusion is unsupported.
Configuration and implementation checks
| Configuration item | Assumed starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Supplier qualification library, single vector model | Check each library's model identifier first, exclude mismatched models, and do not perform cross-model association |
| Search mode | Hybrid retrieval | Entity names and registration numbers are exact fields, and semantic recall can pull in near-identical names |
| Minimum relevance | Assumed starting point 0.35 | Compare cited passages for the same question set at 0.30 and 0.40, and observe whether unrelated entity fragments decrease |
| Citation cap | Assumed starting point 1500 tokens | Compare with 1000 and 2000, and observe whether key sentences in the authorization chain are truncated |
| Result reranking | Enabled | When candidates are numerous, filter with the rerank score, and record the rerank score separately from the vector score |
| Question optimization | Enabled for multi-turn follow-ups | Not enabled for a single clear question; compare cited passages under the two settings |
These are paper checks. No actual retrieval was run, no onboarding was submitted, and no approval or synchronization action occurred.
Three common errors
- Treating vector similarity as entity filtering. Input "Example Supply Chain Management Co., Ltd. authorization" and compare the entity name field in the recalled fragments. If another company's authorization letter is returned, similarity did not perform entity isolation, and the entity field must be checked by hand.
- Treating RRF as a switch that can be turned on. Input the same question and compare the order of cited passages under hybrid retrieval and under semantic retrieval alone. RRF is a fusion step, and no separate switch item exists for it.
- Treating the existence of an authorization letter as onboarding submission eligibility. Input the draft sentence above and compare whether the authorization letter text states a platform onboarding submission right. When the text does not state it, the conclusion is unsupported and the evidence gap is recorded.
How to verify the setup
- Input the business license text and the draft sentence "Example Supply Chain Management Co., Ltd. has an operating status of active," then compare the operating status field in the draft with the source text. The condition is met only when the field and its value match. When the source does not state the field, record it as pending verification.
- Input the authorization chain text and the draft sentence "This authorization is valid until 31 March 2027," then compare the date field. The condition is met only when they match. If the draft says "valid indefinitely," that conflicts with the validity field in the source, and the draft is judged unsupported.
- Input a paper query with no corresponding material and check whether the output citation array is
[], then check whether the downstream path continues on an empty result. Do not claim that a fabricated-rule query necessarily returns an empty array.
All of the above are paper checks. No actual retrieval was executed, no onboarding was submitted, and no approval or synchronization action was produced.