What the documents look like
Marketplace category access happens before a listing is published or a store is onboarded. The core material is the entity qualification and the chain of authorization. A business license, a trademark registration certificate, a brand authorization chain, and an industry permit together form one evidence set that has to be checked item by item. These materials are updated at irregular intervals and resubmitted at each onboarding, so the retrieval side faces text with differing versions, differing periods, and differing field names.
The task of contract and transaction material comparison is to place a draft sentence or a handover request under review alongside existing source material in the knowledge base and compare them on the same entity, the same period, and the same field. The retrieval step is responsible for recalling material that may be relevant; the comparison step is responsible for judging whether fields agree. A retrieval result does not by itself constitute authorization, and it does not confirm that a qualification is valid.
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 using different vector models cannot be associated together. When candidate bases are present, the vector model identifier should be checked first, non-matching bases excluded, and the actual observation recorded. A cross-model retrieval should not be run in order to compare missed recalls.
A single retrieval combines semantic retrieval, full-text retrieval, hybrid retrieval, question optimization, and reranking according to the input content and search parameters, and finally fuses the results into reference content. The optional paths listed in the documentation do not mean that every request executes all of them: image retrieval requires a corresponding image input and model capability, and reranking requires a model and must be enabled by configuration. Reference content is output as an array, and its length can be 0. Even when no content is found, the output link still runs through.
The minimum relevance score is a filtering threshold and takes effect only in semantic retrieval or when result reranking is used. It does not perform permission or entity isolation. A similarity score cannot be used to exclude unrelated businesses or unauthorized records. The retrieval step does not perform authorization judgment; authorization status has to be checked separately in a downstream step.
Hypothetical paper exercise sample: the source material of a brand authorization chain under review reads "Authorizing party A authorized authorized party B on 1 January 2024 to sell a certain category of goods in the East China region, with an authorization period ending 31 December 2025." A draft sentence reads "B's authorization to sell the certain category of goods in the South China region is valid until 31 December 2025." What is to be determined is whether the draft sentence is supported by the source material. The source material's authorized region is East China, the draft sentence states South China, and the two values are mutually exclusive in the authorized region field; the authorization period field agrees in both. This sample is a paper exercise used to illustrate the comparison method and does not represent any real entity or any comparison that has taken place.
Configuration and implementation checks
The table below lists only settings or implementation checks directly relevant to this page's task that are already described in the official documentation. Where a value is needed, it is marked as a starting point for this hypothetical trial, with the reason for choosing it and the object of comparison.
| Configuration item | Starting point for this hypothetical trial | Reason for choosing it and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only one knowledge base that uses the same vector model | Check the vector model identifier of each candidate base first; exclude those whose identifiers do not match; do not associate across models |
| Search mode | Hybrid retrieval | The authorization chain contains proper terms such as entity names and regions; hybrid retrieval uses semantic recall and full-text recall together, then fuses them through RRF |
| Reference upper limit | Starting point set to cover the token volume of one authorization chain material | This value is measured in tokens; compare whether reference content is complete using actual materials of different lengths, then adjust |
| Minimum relevance | Starting point set to a medium value | This value takes effect only in semantic retrieval or result reranking; label relevant and irrelevant samples first, then compare filtered results |
| Question optimization | Enabled for multi-turn follow-up questions containing references | Used for reference resolution and question expansion; compare the original question with the optimized question and check whether the entity and region were altered incorrectly |
| Result reranking | Enabled when candidates are numerous | Reranking performs a second ordering of candidate text and produces a score from 0 to 1; compare ordering and scores before and after reranking |
RRF is the fusion process for hybrid retrieval results and is not configured as an independent switch. The reference upper limit uses tokens as its unit and is unrelated to the thinking budget.
Three common errors
- Treating the "South China region" in the draft sentence as already stated in the source material. The source material's authorized region field is East China, the draft sentence states South China, and the values in the same field are mutually exclusive. This is a contradiction with the source material and should be recorded as unsupported.
- Treating a retrieval hit as confirmed authorization. Returning reference content only means relevant material was recalled. Whether the authorization is valid and whether it covers the target region must be checked field by field against the source material. The retrieval step does not perform authorization judgment.
- Treating an onboarding request under review as already approved. A request in a pending-review state does not demonstrate a lack of operating permission, and it does not constitute approval. Signing, reaching cooperation, and contract effectiveness are three distinct facts. When the material does not state them, the evidence gap is preserved.
How to verify the setup
- Input the source material and the draft sentence from the hypothetical sample, and compare the authorized region field. The source material states East China, the draft sentence states South China, and the values in the same field are mutually exclusive. The observed phenomenon is that the comparison result is marked as contradicting the source material, and that conclusion covers only the region field; the period field is checked separately.
- Input a draft sentence that states only "the authorization period ends 31 December 2025" without stating the authorized region, and compare whether the source material states a region. The source material states East China, the draft sentence does not state a region. The observed phenomenon is that the region field is unsupported and pending verification; it cannot be judged as a negation or a conflict.
- Input a query that contains no qualification keyword at all, and compare the returned reference content array. If the array length is 0, the paper input
[]can be used to check whether the downstream design handles an empty result. An empty result only means no content was recalled this time; it does not mean that no relevant material exists in the knowledge base.