What the documents look like
Platform refund and return handling runs order by order rather than as a single live-sale settlement. A refund may arise without any promise made during one broadcast session, so the material set is assembled per order and refreshed per order and per day. Four document types carry the comparison: the refund application, the return logistics note, the quality inspection photo, and the refund approval record. These are held across the after-sales backend, customer service chat, WMS receiving, and payment refund records. The comparison task places the application content, the return parcel status, the inspection conclusion, and the approval record of one order side by side, then checks whether the subject, the object, the amount, and the status stay consistent. Retrieval and recall supply the quoted content that the comparison reads; retrieval does not decide responsibility and does not execute a refund.
Constraints at the retrieval step
Material is generated per order, so search terms mix exact fields such as order number, logistics number, and approval status with natural-language descriptions such as return reason and inspection conclusion. The knowledge base search node can associate one or more knowledge bases that use the same vector model; bases built on different vector models cannot be associated together. The quoted-content output is an array whose length may be 0, and the output path still completes when nothing is found.
The reference cap is measured in tokens, not in top k, because in a mixed knowledge base the chunk lengths differ widely and a token cap is steadier. The minimum relevance value is a filter threshold that applies only under semantic retrieval or when result reordering is used; it is not a final sort rule and it does not act as a permission or entity filter. A relevance score is not a category, entity, or permission filter, and retrieval is not authorization. Approval and refund actions require downstream implementation. Question optimization performs coreference resolution and question expansion before retrieval, which adds a model call; it can help multi-turn retrieval and it can also be turned off when the current question is already clear or when response time matters more. Result reordering requires a reordering model and must be enabled by configuration; the reordered score is a 0-1 value that can be used for filtering. Hybrid retrieval merges semantic and full-text results through RRF, so the final quoted content is not necessarily ordered by a single vector similarity.
Hypothetical paper exercise sample. The refund application text reads: "Order A-2201, buyer applied for a return on 4 March, return reason recorded as size mismatch, refund amount requested 199." The return logistics note reads: "Return parcel number T-8801, received by WMS on 6 March, receiving status recorded as warehoused pending inspection." The inspection and approval text reads: "Inspection conclusion: appearance intact, accessories complete; refund approval status is pending review." The draft sentence to be judged is: "This order has completed its refund." The comparison must confirm whether the requested amount and the approved amount agree in the same field, whether the return parcel has been signed for, and whether the approval status has moved from pending review to approved. Pending review does not constitute a completed refund.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only one refund material base that uses a single vector model | Verify the vector model identifier of each base, exclude those with a different identifier, and do not run a cross-model retrieval |
| Search mode | Compare semantic retrieval and full-text retrieval in parallel | For order-number questions observe full-text hits; for return-reason questions observe semantic hits; record each result |
| Reference cap | Set to cover one complete refund application plus one logistics record in tokens | Test with actual material of different lengths to see whether content is truncated, then adjust |
| Minimum relevance | Label relevant and irrelevant samples first, then choose a starting value | This value applies only under semantic retrieval or result reordering; compare filtering against the labelled samples |
| Question optimization | Enable for multi-turn follow-up scenarios, disable for a single clear question, and compare | Compare the original question with the optimized question and check whether the subject was altered |
| Result reordering | Enable when candidates are numerous and compare reordered scores | Reordering requires the model and must be enabled by configuration; the score is used for filtering |
Three common errors
- Treating a retrieval hit as evidence that a refund has completed. In the sample the approval status is pending review; recalling that record shows the material exists, and it does not show the refund action was executed.
- Associating knowledge bases that use different vector models and then comparing missed recall. The correct step is to verify the vector model identifier first, exclude those that do not match, and record what is actually observed without inferring a runtime result.
- Treating the minimum relevance value as a permission or entity filter. The value is a relevance threshold; it cannot exclude unrelated businesses or unauthorized records, and retrieval by itself is not authorization.
How to verify the setup
- Input the draft sentence from the sample, "This order has completed its refund," and compare the value of the refund approval status field. The sentence is supported only when that field is explicitly approved and the amount agrees with the requested amount in the same field. The current value is pending review, so the sentence is not supported.
- Input the return logistics note text and compare the receiving status field. The return leg is supported only when that field is explicitly signed for and the inspection conclusion has been issued. In the sample the parcel is warehoused pending inspection and the inspection conclusion has been issued, so the link between signing and inspection still needs to be checked.
- Input a paper retrieval input that is explicitly empty, and check the downstream design for handling an empty result array. Empty-result handling is satisfied only when the output path still completes at length 0 and the downstream has a corresponding branch. No claim is made that a fictional rule-based retrieval must return an empty array.
This page covers only the retrieval and recall step for document comparison in platform refund and return handling. Whether material, an operation, or a request is authorized at handover cannot be settled by retrieval or by a form.