What the documents look like
Marketplace refund and return handling occurs order by order. It does not depend on a promise made during a single livestream session. The working materials come from the platform after-sales backend, customer service chat, WMS receiving, and payment refund records. In day-to-day operation these appear as refund applications, return logistics notes, quality inspection photos, and refund approvals. Operating reviews and exception reviews need to place these materials under one order subject and compare them together.
Retrieval therefore faces two clearly different input shapes. One shape is natural-language description of a return reason. The other shape is precise fields such as order number, logistics number, and approval status. Materials update per order and daily. A retrieval configuration needs to hold both semantic questions and exact identifier matching.
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. The smallest search unit is the knowledge base. A collection is only a management grouping and does not change search behavior.
Semantic retrieval depends on vector distance. Full-text retrieval depends on keyword hits. Hybrid retrieval merges the two paths through RRF. Result reranking uses a ReRank model for a second ordering pass. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. The reference limit is measured in tokens. Question optimization performs coreference resolution and question expansion, and it adds one model call before the actual retrieval. Reference content is output as an array and its length can be 0, so the output path still runs when nothing is retrieved. Retrieval and similarity scores do not carry permission or entity isolation, and they do not constitute authorization.
Hypothetical drill sample: the source material reads "Refund application A-001, order subject Hypothetical Shop Jia, application date 10 March 2025, refund reason damaged goods, current status pending review." A second line reads "Return logistics note B-002 linked to order subject Hypothetical Shop Jia, signed date 12 March 2025, receiving warehouse Hypothetical Warehouse." A third line reads "Quality inspection photo C-003 linked to order subject Hypothetical Shop Jia, taken 13 March 2025, shows outer packaging damage." The draft sentence to be judged is "The order of Hypothetical Shop Jia has completed refund approval." The draft sentence shares the same subject and the same order scope as the source material, but the material records only a pending review status and does not record a completed approval.
Configuration and implementation checks
The table below lists documented settings or implementation checks relevant to this page's main task, tied to the sample above. Any numeric value is a hypothetical starting point for this drill and states what it is compared against. It is not an official default.
| Configuration item | Hypothetical drill starting point | Comparison and check method |
|---|---|---|
| Associated knowledge bases | Associate only the knowledge base holding refund applications, return logistics notes, quality inspection photos, and refund approvals | Check that the vector model identifier is the same across bases; exclude any base whose identifier differs and do not run a cross-model association |
| Search mode | Use semantic retrieval and full-text retrieval together | Test with the sample natural-language question and the sample order number question; observe how the two recall paths differ |
| Reference limit | Hypothetical starting point of 800 tokens | Test with a short and a long refund application and inspection note; compare content completeness against a longer limit |
| Minimum relevance | Hypothetical starting point of 0.4 | Mark relevant and irrelevant fragments in the sample first, then compare the recalled set at different thresholds |
| Question optimization | Enabled for follow-up turns | Compare the original question with the optimized question and check whether the order subject and status were changed incorrectly |
| Result reranking | Enabled when hybrid retrieval returns many candidates | Check whether a rerank model is present and enabled by configuration; compare ordering before and after reranking |
Three common errors
- Treating a pending review status as a completed approval. The input is the sample material line "current status pending review." The comparison object is the draft sentence "has completed refund approval." The observable result is that the material does not record a completed approval, so the draft sentence is not supported.
- Treating a retrieval hit as authorization. The input is the refund application fragment returned by retrieval. The comparison object is the actual authorization state. The observable result is that the retrieval result contains no authorization conclusion, so no refund or approval action can be executed on that basis.
- Forcing an association across knowledge bases with different vector models. The input is two bases with different vector model identifiers. The comparison object is the node's association requirement. The observable result is that the identifiers do not match, so the base should be excluded and the observation recorded, without inferring how the association would behave at runtime.
How to verify the setup
- Check one: input the full sample material lines and the draft sentence "The order of Hypothetical Shop Jia has completed refund approval." Compare the status field recorded in the material with the wording of the draft sentence. The observable result is that the material records only pending review, so the draft sentence is not supported and the evidence gap must be recorded.
- Check two: input the sample order number and the sample return reason, and run semantic retrieval and full-text retrieval separately. Compare the recalled content from the two paths. The observable result is that the identifier question returns more stable hits under full-text retrieval and the natural-language question returns closer content under semantic retrieval; the difference between the two must be recorded.
- Check three: input an explicit paper empty result and check whether the downstream design accepts an empty array. The observable result is that the output path still runs. No claim is made that a fictional rule retrieval necessarily returns an empty array.