What the documents look like
Platform refund and return handling occurs order by order. An after-sales request does not necessarily connect to a promise made in a single livestream session, which distinguishes this category from centralized livestream after-sales processing. The daily materials for citation tracing are the refund application, the return logistics waybill, quality inspection photos, and the refund approval record. Customer service chat logs, WMS receipt records, and payment refund records supply additional status information. These materials sit in different systems and different stages, so tracing a citation requires confirming the material version and its valid scope for each order.
Materials update per order and per day. A refund application and a return logistics waybill for the same order can belong to different points in time. When an answer cites a material, the currently effective version must be checked. A refund application that has been submitted, a waybill that shows warehouse receipt, and a quality inspection photo are separate records with separate evidentiary weight. None of them, on its own, states an approval conclusion or a payment outcome.
An answer in order fulfillment and after-sales question handling that cites materials must point to a specific document. The answer text and the cited source text are compared item by item. Where the source does not record a field, that field remains unknown.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that use the same vector model. Its output is an array of citations, and the array length can be zero. When the citation array is empty, the search path still completes; an empty array does not mean the workflow has errored out or terminated. The answer step that receives the citation input must check, one by one, how the draft answer corresponds to the cited source text.
Retrieval results do not carry authorization. Similarity scores, the minimum relevance threshold, and the citation token limit are retrieval parameters. They do not replace business review. Whether a refund approval passes and how return responsibility is assigned must be decided by the actual review actions in the after-sales process. Values collected by retrieval or by a form do not constitute authorized execution.
A relevance score is not a category, entity, or permission filter. A similarity setting does not exclude unrelated businesses or unauthorized records. Retrieval is not authorization. Forms collect values; business verification, approval, database writes, and external synchronization require their stated downstream implementation.
Hypothetical drill sample: the source material reads, "Order A's refund application was submitted on March 1. The return logistics waybill shows the warehouse signed for receipt on March 3. The quality inspection photo records that the outer packaging was damaged." One draft sentence to be assessed reads, "Order A's refund has been approved, and the payment will be returned on March 5." Comparing against the source, the material records only the application submission, the warehouse receipt, and the inspection record. It does not record an approval conclusion or a refund arrival date. The phrases "has been approved" and "will be returned on March 5" are not supported by the source text and remain items to be verified. They cannot be output as facts that have occurred.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this page | How to check |
|---|---|---|
| Associated knowledge bases | Select one refund after-sales knowledge base using a single vector model | Check that the selected knowledge base identifier and the vector model match; exclude knowledge bases with a different vector model |
| Search mode | Hybrid search | Observe whether semantic recall and full-text recall both participate, and whether results are fused through RRF |
| Citation limit | This hypothetical trial starts at 3000 tokens | This value is only a trial starting point; compare it against actual recalled content and adjust |
| Minimum relevance | This hypothetical trial starts at 0.3 | This value takes effect only under semantic search or result reranking; compare cited text against the source item by item |
| Question optimization | Enabled for multi-turn follow-up scenarios | Observe whether the completed question matches the subject of the current order material |
| Result reranking | Enabled when candidate results are numerous | Check the rerank score against source relevance; the score is not used for permission filtering |
These configuration items cover only the retrieval and citation step. Refund approval, payment return, and responsibility determination are downstream business actions. They must be implemented by the after-sales system or by human review. The search node does not execute these actions.
Three common errors
- Treating an empty citation array as a workflow error. When the input is a clearly written empty result, check whether the downstream answer step is designed to accept an empty citation, and observe whether the output path continues. Do not assert that a retrieval rule necessarily returns an empty array.
- Treating the minimum relevance threshold or a similarity score as a permission or entity filter. During checking, observe that the score is used only for relevance ordering, not for excluding unauthorized orders or unrelated merchants. Authorization decisions must connect to a separate business review.
- Treating a pending review or a submitted application as an approved refund. In the hypothetical sample, the refund application has been submitted and the return has been received, while the approval conclusion is not recorded. Compare the draft sentence against the source and record the unrecorded field as pending verification.
How to verify the setup
- Input the hypothetical sample's source text and the draft sentence. Compare whether "has been approved" and "will be returned on March 5" in the draft sentence have corresponding records in the source. The source records only the application submission, the warehouse receipt, and the inspection record. The draft sentence's related statements are not supported, and the check result should be pending verification, not a determination that approval or refund has occurred.
- Input the refund application and the return logistics waybill for the same order. Compare whether the order identifier and the time range in the two materials match. If the waybill receipt date is later than the application date and the order identifiers are the same, check whether the cited content points to that order and observe whether the answer corresponds to the source item by item.
- Input an order number that does not appear in the knowledge base. Observe the citation array output and the downstream acceptance. If the citation array length is zero, check whether the downstream answer step continues to execute as designed for an empty citation, and record the actual observation. Do not presuppose that retrieval necessarily returns an empty array or necessarily finds a match.
The primary task of this page is citation tracing for after-sales questions in platform refund and return handling. The page does not decide whether a refund is approved, whether payment is returned, or how responsibility is assigned. Those decisions remain with the after-sales review process.