What the documents look like
Refund-only disputes differ from ordinary return-and-refund cases in one structural way: there is often no returned parcel entering the warehouse. The evidence therefore concentrates on dispatch and signed receipt, together with the value of the goods. The document set for this category includes refund-only applications, dispatch evidence, signed-receipt proof, and appeal materials. These come from the platform after-sales backend, carrier tracking, customer-service chat, and warehouse outbound records. Materials are produced per order, and the brief describes their update frequency as per order or daily.
For operations handover, the decisive question is not whether a document can be found. It is what authorization state the material, the operation, and the request each occupy at the moment of handover. Retrieval is not authorization. A retrieved appeal material does not mean that appeal has been approved. A retrieved application does not mean the application has been granted.
Constraints at the retrieval step
The knowledge-base search node can associate one or more knowledge bases that use the same vector model. Bases using different vector models cannot be associated together. The identifier of each base must be checked first, and non-matching bases excluded. Cross-model retrieval is not performed in order to compare missed recalls.
Referenced content is output as an array, and its length can be zero. Even when no content is found, the output path still completes. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It is a filtering threshold, not the final ranking rule, and it does not perform permission or entity isolation. The same-vector-model requirement for associated bases does not extend to external systems.
Hypothetical drill sample: an after-sales supervisor at a shop submits a handover request proposing to transfer handling authority for "refund-only application A-0715" to an appeal specialist. The material text supplied for comparison reads: "Application A-0715 was created on 15 July and its current status is pending review. Dispatch evidence and signed-receipt proof have been uploaded to the after-sales backend. Appeal materials have not yet been submitted." The draft sentence to be judged reads: "A-0715 has been approved, refund-only granted." The subject, object, and status needed for judgment are all stated. Pending review does not demonstrate a lack of operating authority, and it does not equal approval. The draft sentence and the source status field are mutually exclusive, so the draft sentence is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this page | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate one base containing refund-only applications, dispatch evidence, signed-receipt proof, and appeal materials | Check whether the vector-model identifiers of the bases are the same; exclude those that differ |
| Search mode | First compare semantic retrieval, full-text retrieval, and hybrid retrieval separately against the test question | No single mode is presumed suitable for all materials; observe by question type |
| Reference cap | Measured in tokens; start at a value that can hold a complete application form | Test with materials of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Label relevant and irrelevant samples first, then set a starting value between 0 and 1 | Takes effect only under semantic retrieval or result reranking; compare against the labelled results |
| Question optimization | Enable for continuous follow-up questions that contain references; compare the original question with the optimized question | Check whether the subject has been altered incorrectly; clear questions can be left off to save time |
| Result reranking | Enable when candidates are numerous and a rerank model is available | Rerank score runs from 0 to 1 and filters by score; confirm the model is configured |
Three common errors
- Treating a retrieval hit as authorized execution. Input the draft sentence "A-0715 has been approved" and compare it with the source status field "pending review." If the draft sentence is written into the handover record as an approval, authority has been exceeded. Executable check: read the source status field, compare it character by character with the status declared in the draft sentence, and mark it unsupported when the two do not match.
- Treating pending review as a lack of operating authority. Input the authority-transfer item in the handover request and compare whether the source text states a field granting or revoking authority. When the source does not state it, record it as pending verification rather than as a negative finding.
- Associating bases that use different vector models. Input the vector-model identifiers of the two bases and compare whether they match. Exclude the base whose identifier differs, and do not run cross-model retrieval to infer missed recall or redundancy.
How to verify the setup
- Against the sample source text "Application A-0715 was created on 15 July and its current status is pending review," input the draft sentence "A-0715 has been approved, refund-only granted." Compare the source status field with the status declared in the draft sentence. The observed result is that the two are mutually exclusive, so the draft sentence is judged unsupported and must not enter the handover record.
- Against the authority-transfer item in the sample handover request, input the source text "Appeal materials have not yet been submitted." Compare whether the source states a field granting, revoking, or bringing authority into effect. The observed result is that it does not state one, so the item is recorded as pending verification and the evidence gap is preserved.
- Against the associated knowledge bases, input the vector-model identifiers of the two bases. Compare whether the identifiers match. When they do not match, exclude that base, do not run cross-model retrieval, and record the actual observation. Empty-result handling can be checked against the explicit paper input
[]for downstream design; no claim is made that a fictional rule search must return[].