What the documents look like
Platform refund and return handling happens order by order. It does not depend on a promise made during one live stream. In store service and internal training, support agents and after-sales specialists work with four recurring materials: the refund application, the return logistics document, the quality inspection photo, and the refund approval. These materials are produced per order and are spread across the platform after-sales backend, agent chat, warehouse receiving, and payment refund records. The citation tracing step has one job: when a store employee hands off a case or trains a colleague, that person can point to which material, which document version, and which passage supports a stated conclusion, and can separate a retrieval hit from a business authorization.
Constraints at the citation tracing step
The knowledge base search node outputs cited content as an array, and the length may be zero. An empty array still passes through the output link. The citation limit is controlled in tokens. The minimum relevance value takes effect only under semantic retrieval or result reranking. Neither setting performs permission filtering or entity isolation. Knowledge base association requires the same vector model; libraries with different vector models cannot be associated together. Retrieval itself does not constitute authorization. Whether a refund approval has passed and how return responsibility is assigned must be implemented in downstream business steps.
Hypothetical drill sample: the store training material reads, "The refund application for Order A was submitted by the agent on March 12. The return logistics document shows the warehouse received the goods on March 14. The quality inspection photo records outer packaging damage. The refund approval status is pending review." The draft sentence to be judged reads, "This order has completed refund approval, so the customer can be told the refund has passed." The comparison must hold the subject (Order A), the object (refund approval), the status (pending review and passed are mutually exclusive values in the same field), and the date range (March 12 to March 14). Pending review does not prove approval passed, and it does not prove a lack of operating permission.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Comparison method and reason |
|---|---|---|
| Associated knowledge base | Select one refund after-sales knowledge base that has completed training | Check whether the vector model identifier matches; exclude libraries with different models and do not run joint retrieval |
| Search mode | Hybrid retrieval | Refund order numbers and logistics numbers need full-text recall; natural language questions need semantic recall |
| Citation limit | Hypothetical starting point of 3000 tokens | This is only a paper comparison starting point for this drill, compared against 1500 tokens to observe the change in the number of cited items; it is not an official default |
| Minimum relevance | Hypothetical starting point of 0.3 | This is only a paper comparison starting point for this drill, compared against 0.5 to observe the number of items remaining after filtering; it is not an official default |
| Result reranking | Enabled according to settings | A rerank model must already be configured; when it is not configured, this item does not execute |
| Citation input handoff | Pass the citation array to the answering step | Check each answer sentence against the cited original text; do not substitute answer fluency for source evidence |
Three common errors
- Treating a retrieval hit as a passed approval. Input the draft sentence and the training material original text, compare the refund approval field value, and note that pending review and passed are mutually exclusive values in the same field. When the material states only pending review, the draft sentence is not supported.
- Treating an empty citation array as a workflow error. Input a paper empty array, compare whether the downstream node still passes through, and note that the node output allows a length of zero. An empty array does not mean the workflow has terminated.
- Associating libraries with different vector models. Input the vector model identifiers of two libraries, compare whether the identifiers match, and exclude them when they differ. Do not run cross-model retrieval to compare missed recalls.
How to verify the setup
- Input the training material original text and the draft sentence. Compare the refund approval status field. When the material states pending review and the draft sentence claims passed, the draft sentence is not supported, and the evidence gap is recorded.
- Input a paper empty array. Compare the downstream handoff design. When the output link passes through and the downstream does not interpret the empty array as an approval result, the empty-result handoff condition is met.
- Input the vector model identifiers of two knowledge bases. Compare whether the identifiers match. Joint association is allowed only when the identifiers match. When they differ, exclude them and record the actual observation.