What the documents look like
Marketplace refund and return handling happens order by order across the platform after-sales back office, customer service chat, warehouse receiving records, and payment refund records. These materials form independently and do not attach to a single livestream commitment. A refund application, a return logistics document, a quality inspection photo, and a refund approval can each carry a different version and timestamp, so drafting product and marketing copy that quotes them requires confirming which version and which retrieval scope apply. Materials update per order and per day. When a draft cites the same order field from two documents created at different points, the values must be compared field by field rather than assumed consistent.
Constraints at the citation tracing step
The knowledge base search node outputs a citation content array, and its length may be zero. An empty array means the output path still runs; it does not mean the workflow has errored or terminated. The citation limit controls the maximum tokens of content cited per search. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used, and it is a filter threshold rather than a permission or entity isolation control. Knowledge bases can be associated together only when they share the same vector model; bases with different vector models cannot be associated together, so check the identifier first and exclude those that do not match. Hybrid retrieval fuses multiple result paths through RRF, which is a fusion process rather than a switch that can be turned on. Reranking requires a ReRank model and must be enabled by configuration. Image retrieval requires image input and model capability. Neither implies that every request executes that path.
Hypothetical drill sample: a platform after-sales knowledge base holds one refund application original text: "Order A-2201, buyer applied for return on March 8, 2025, reason is that the item does not match the page description, application status is pending review." A second quality inspection photo description original text reads: "Order A-2201, inspection date March 12, 2025, inspection conclusion is that the appearance is intact, no mismatch with the description found." The draft sentence to be judged is: "Order A-2201 has confirmed the item does not match the description, and the refund has been approved." Comparison checks the draft sentence against both original texts: the refund application status is pending review, and the inspection conclusion is that the appearance is intact. The claims "has confirmed the mismatch" and "the refund has been approved" appear in neither original text, so they are unsupported.
Configuration and implementation checks
| Configuration or check item | Purpose on this page | Value and basis |
|---|---|---|
| Associated knowledge bases | Limit the scope searchable for drafting citations | Select one or more knowledge bases with the same vector model; different vector models cannot be associated together |
| Citation limit | Control the maximum tokens of content cited per search | For this hypothetical trial the starting value is 2000 tokens, used to compare whether citations cover the sample original texts; the basis is the need to hold both texts with margin, compared against 1000 tokens to observe truncation |
| Minimum relevance | Filter low-relevance results | Takes effect only in semantic retrieval or result reranking; for this hypothetical trial the starting value is 0.5, compared against 0.3 to observe whether the sample original texts are still recalled |
| Search mode | Determine the recall path | Semantic retrieval, full-text retrieval, or hybrid retrieval; identifier-type fields can use full-text retrieval |
| Result reranking | Reorder candidates a second time | Requires a ReRank model and must be enabled by configuration; when not enabled, no rerank score is produced |
| Citation content output | Carry retrieval results to the answer step | Output is an array whose length may be zero; an empty array still requires downstream handling by design |
Three common errors
- Writing a pending-review status as approved. In the sample, the refund application status is pending review, while the draft sentence says "the refund has been approved." Comparing the same order and the same field, the two values are mutually exclusive, which is a contradiction with the original text.
- Writing an inspection conclusion of intact appearance as a confirmed mismatch. In the sample, the inspection original text concludes that the appearance is intact, while the draft sentence says "has confirmed the item does not match the description." The same order and the same inspection field hold mutually exclusive values, which is a contradiction with the original text.
- Associating knowledge bases with different vector models and then searching directly. The documentation states only that the same vector model is required for joint association; it does not state the runtime result of associating across models. The correct step is to check the identifier first, exclude those that do not match, and record what is actually observed without inferring missed recalls or redundancy.
How to verify the setup
- Input the sample refund application original text and the draft sentence "Order A-2201 has confirmed the item does not match the description, and the refund has been approved." Compare the draft sentence against the application status field in the original text, and observe whether the draft sentence contains "has confirmed the mismatch" and "has been approved," which the original text does not state. If they appear, the judgment is unsupported.
- Input the sample quality inspection photo description original text and the same draft sentence. Compare the inspection conclusion field, and observe whether the draft sentence is mutually exclusive with "the appearance is intact." If it is mutually exclusive, the judgment is a contradiction with the original text.
- Input an explicit paper empty array [] to the downstream receiving step, and compare whether the downstream design continues processing under an empty citation. If the downstream runs by design without error, the empty-result receiving condition is satisfied. Whether retrieval actually returns an empty array depends on the actual data and parameters, and is not asserted in advance.