What the documents look like
Marketplace refund and return handling occurs per order on a continuing basis, and it does not necessarily connect to a promise made during a single livestream session. The materials involved are refund applications, return logistics waybills, quality inspection photographs, and refund approvals. These materials come from the after-sales backend, customer service chat, WMS receiving, and payment refund records.
For operating reports and exception reviews, the question is which material and which approval a given refund conclusion relied on at the time, and whether that material can still be retrieved during the review. Every amount, responsibility determination, and refund status in a report needs to point back to a specific source. The citation tracing step handles the correspondence between retrieved source text and report conclusions. It does not handle whether the refund amount itself was reasonable or whether the responsibility determination was correct.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that share the same vector model. Its output is an array of citations, and the length may be zero. When the array is empty, the output path still completes. The citation array is passed to a downstream answer or reporting step, where each answer is checked against the cited source text. Retrieval by itself does not constitute authorization, and it does not constitute execution of a refund, an approval, or a database write.
Hypothetical drill sample: a marketplace after-sales team prepares a March refund exception review. The source text consists of three sentences. Sentence one: "Order A-2291, refund application submitted 2026-03-05, requested amount 128 yuan, status pending review." Sentence two: "Return waybill number SF-7731, signed for 2026-03-09, receiving warehouse East China warehouse." Sentence three: "Inspection photo Q-0417 records outer packaging damage to the goods, taken 2026-03-10." The draft sentence to be assessed is: "Order A-2291 has completed refund approval, and the refund amount of 128 yuan has been returned via the original payment channel."
Authorization boundary in this sample: the materials above are used only to check citation sources. Whether refund approval was completed and whether funds were returned are not stated in the materials and remain unsupported items. The material subject is order A-2291, the objects are the refund application and the return logistics, the status field is pending review, and the effective scope is March 2026.
Configuration and implementation checks
The table below lists retrieval settings and manual verification items relevant to this feature and primary task, based on the sample above. The values are hypothetical starting points for this exercise, used to compare against another value and observe differences. They are not official default values.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select only the refund after-sales material base | First verify that the selected base identifiers and vector models match; knowledge bases with different vector models cannot be associated together |
| Search mode | Hybrid search | Compare with semantic search alone and observe whether identifier fields such as order numbers are hit |
| Citation limit | Hypothetical starting point 3000 tokens | Compare with 1500 tokens and observe whether all three sample sentences enter the citation |
| Minimum relevance | Hypothetical starting point 0.4 | Compare with 0.2 and observe whether the pending-review status sentence is filtered out |
| Result reranking | Enabled per settings | Confirm the rerank model is available and compare citation ordering before and after enabling |
| Citation content output | Check source text line by line | Compare the citation array against the source sentences and record the empty-array case |
Three common errors
- Treating retrieval output as an approval result. In the sample, the refund application status is pending review. Retrieving that sentence only shows the material was recalled; it does not show refund approval was completed. Pending review and approved are mutually exclusive values of the same field. The material gives only pending review, while the draft sentence states approval was completed, which is unsupported.
- Treating a zero-length citation array as a process failure. The search node output citation array may have length zero, and the output path still completes when the array is empty. Use the explicit paper input [] to check the downstream design, record how the downstream handles empty citations, and do not infer that retrieval must return empty.
- Treating a relevance threshold as a permission or entity filter. Minimum relevance is a filtering threshold and does not serve to exclude unrelated business or unauthorized records. Retrieval is not authorization. Knowledge bases with different vector models cannot be associated together; check identifiers only and exclude those that do not match.
How to verify the setup
- Input the three sample source sentences and the draft sentence. Compare "has completed refund approval" in the draft sentence with "status pending review" in the material. The material does not state that approval was completed, so the item is judged unsupported. Preserve the evidence gap and do not supply authorization or effective clauses.
- Input order number A-2291 and compare the citation arrays from hybrid search and semantic search alone. Observe whether the order number appears in the cited source text and record the actual observation. Do not guarantee a first-position hit or full recall.
- Input the paper empty citation []. Compare how the downstream reporting step is designed to handle an empty array. Observe whether the downstream records a gap or takes an exception branch. Do not claim that a fictional rule retrieval must return [], and do not infer automatic business actions.