What the documents look like
Cross-border platform settlement and foreign exchange work concerns the collection of funds from orders that have already been completed. It does not handle fraud chargebacks, which belong to payment risk control. The working materials are platform settlement statements, exchange rate tables, service fee invoices, and receipt reconciliation tables. These update weekly or monthly.
A settlement statement records order deductions and payout batches. An exchange rate table records the conversion time point and the conversion basis. A service fee invoice records which party bears a fee. A reconciliation table records the matching result between receipts and orders. During an operations handover, these same materials also carry the boundary of authority: who may view the settlement data of a given shop for a given period, who may raise a review request for a reconciliation difference, and who may confirm that a handover is complete. The retrieval step organizes these materials by knowledge base so that the person taking over can find the settlement basis that matches that person's own authority scope. Retrieval output is not an authorization credential.
Constraints at the retrieval step
A knowledge base search node can associate one or more knowledge bases that use the same vector model. The search mode is chosen per test question among semantic retrieval, full-text retrieval, and hybrid retrieval. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result re-ranking is used. Question optimization performs coreference resolution and question expansion. These settings shape how recalled content is organized. They do not decide whether a given settlement record falls inside the current operator's authorization scope.
Retrieval output is an array of citations, and its length can be zero, so the empty-result path still runs through. Hybrid retrieval merges multiple result paths through RRF. Re-ranking requires a model and is enabled according to settings. Image retrieval requires image input and the corresponding model capability. The overall retrieval flow combines text, images, semantic recall, full-text recall, question optimization, and re-ranking, then fuses the multiple paths into citation content. The optional paths listed in the flow diagram do not all execute on every request.
Hypothetical paper exercise: a cross-border seller's finance group prepares to hand over the July settlement reconciliation of one shop to a colleague. The source sentences supplied for checking are three. "The shop A July platform settlement statement records payout batch 202407-A, total deductions of USD 1,200, and status paid." "The shop A July receipt reconciliation table records USD 1,180 received, a difference of USD 20, and the difference reason field is empty." "The handover application form records: request to transfer shop A July settlement reconciliation permission to colleague B, approval status pending review." One draft sentence to be judged is: "Shop A July settlement has been reconciled, the USD 20 difference has been confirmed as caused by a service fee, and the permission has been transferred to colleague B." This draft sentence can be judged only after item-by-item comparison with the three source sentences. The handover application form is pending review, which does not constitute a conclusion that permission has been transferred. The difference reason field is empty, which does not constitute a conclusion that the difference has been confirmed as a service fee.
Configuration and implementation checks
The table below lists, around the exercise above, the documented settings or implementation checks relevant to this feature and the operations handover task. The values are starting points for this hypothetical trial, used to compare recall behavior under different values. They are not official default values.
| Configuration item | Starting point for this hypothetical trial and reason | Comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the knowledge base holding the shop A settlement statement, reconciliation table, and handover application form; confirm these bases use the same vector model | Check the vector model identifier of each base; exclude any base whose identifier differs from the same association |
| Search mode | Start with hybrid retrieval, because the exercise contains batch number 202407-A and a natural-language difference description | Run the same question under semantic, full-text, and hybrid retrieval; compare whether returned citations cover both the batch number and the difference description |
| Citation limit | Set the starting point to a token amount covering both the settlement statement passage and the reconciliation table passage | Test with actual materials of different lengths to see whether citation content is truncated, then adjust the limit |
| Minimum relevance | First label citations relevant to the handover check and citations that are not | Under semantic retrieval or result re-ranking, compare how threshold changes select between the two labeled groups |
| Question optimization | Enable for follow-up questions containing references such as "that batch" or "the above difference" | Compare the original question with the optimized question, and check whether the subject has been altered incorrectly |
| Result re-ranking | Use when a re-ranking model is available and enabled according to settings | Compare candidate citation ordering before and after enabling, and observe position changes of relevant citations |
Three common errors
- Treating a recalled settlement statement citation as evidence that permission has been transferred. In the exercise, the handover application form has approval status pending review. A retrieval hit on that form shows only that the material was recalled; the authorization state still requires checking against the approval process.
- Reading an empty difference reason field as a confirmed difference. The reconciliation table source has an empty difference reason field, which is not stated. The sentence "the difference has been confirmed as caused by a service fee" cannot be written from it.
- Forcing knowledge bases with different vector models into one association. Before associating, check each base's vector model identifier, exclude bases whose identifiers differ, and record the identifiers actually observed.
How to verify the setup
- Input the three source sentences and the draft sentence. Compare the three items "has been reconciled," "the difference has been confirmed as caused by a service fee," and "the permission has been transferred to colleague B" against the source. The source states only the payout batch, deduction amount, received amount, difference amount, an empty difference reason field, and approval status pending review. Observable result: none of the three items is supported by the source, so the draft sentence is marked as pending verification, and the difference reason and permission state remain evidence gaps.
- Input the question "what is the difference between deductions and receipts for shop A July payout batch 202407-A," and under hybrid retrieval compare whether returned citations cover both the settlement statement deduction amount and the reconciliation table difference amount. Observable result: the condition is met when citation content contains values from both passages; when only one passage is hit, record partial coverage and do not judge retrieval as failed.
- Input a paper query containing no relevant material, and check the downstream design for handling empty citations. Citations are output as an array whose length can be zero, and the empty-result path still runs through. Observable result: the condition is met when the downstream continues on an empty array; if the downstream depends on non-empty citations and defines no empty-array branch, record it as a design gap, and do not claim the query necessarily returns an empty array.