What the documents look like
Settlement and foreign-exchange work covers funds collection for orders that have already been completed. It does not handle fraud chargebacks, which belong to cross-border payment risk control. The materials involved are platform settlement statements, exchange-rate tables, service-fee invoices and incoming-payment reconciliation sheets. These materials are updated weekly or monthly, and the fee deductions and exchange-rate differences for a single order must be checked line by line against the original documents.
The parties doing this work are cross-border seller finance staff, tax accountants and platform operations staff. Errors here land directly on profit calculation, tax filing and cash planning. In this category, citation tracing means that when an answer about order fulfilment or an after-sales question is given, the original text the conclusion relies on is supplied together with it, and the version and validity scope of that text are stated. Settlement statements and exchange-rate tables are revised over time, and the same order may correspond to different rate tables in different months, so a citation must point back to that specific version of the material.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that use the same vector model, and it outputs citations as an array whose length may be zero. A zero-length output means the downstream chain still runs even when nothing was found. This zero-length output is the premise for the downstream design: the answering step needs to handle empty citations and state honestly that no material was retrieved.
Retrieval results are fused across semantic search, full-text search, hybrid search, question optimisation and result re-ranking. RRF is the process that merges results after hybrid search; it is not a switch that can be turned on by itself. The minimum relevance value is a filter threshold that takes effect only in semantic search or when result re-ranking is used, and it does not perform permission or entity isolation. Retrieval by itself does not constitute authorisation. Whether a given settlement statement or reconciliation sheet may be handed to the requester depends on the access scope of the knowledge base holding that material and on the outcome of human review.
Image retrieval requires an image in the input and a model with the corresponding capability, and re-ranking requires a configured re-ranking model that is enabled in the settings. These paths do not run on every request. Knowledge bases using different vector models cannot be linked together; when checking, confirm the vector model identifier first and exclude those that do not match, then record what was actually observed.
Hypothetical drill sample
Material text for checking (fictional):
Material one, a platform settlement statement (March 2025 period, number J-2025-03, entity: Store A): amount due this period 100,000 yuan, platform service fee 2,000 yuan, deduction details in the attached schedule.
Material two, exchange-rate table (March 2025 version, applicable period 2025-03-01 to 2025-03-31, entity: Store A): USD to CNY central parity rate 7.10.
Material three, incoming-payment reconciliation sheet (March 2025, entity: Store A): bank credit amount 98,000 yuan, credit date 2025-04-05.
Draft sentence to be judged:
"Store A's March 2025 settlement of 100,000 yuan arrived in March 2025."
The entity and period in this draft sentence match the materials, but material three states a credit date of 2025-04-05, while the draft sentence says the money arrived in March 2025. This is a mutually exclusive pair of values for the same field. The materials do not state the authorised disbursement status of this payment, nor whether review has been completed, so that part remains to be verified.
Configuration and implementation checks
| Configuration item | Starting point for this hypothetical trial | Reason for the choice and comparison object |
|---|---|---|
| Linked knowledge bases | Include the settlement statement base and the exchange-rate table base | Both bases must use the same vector model; check the identifier first, exclude non-matching ones, and do not run cross-model retrieval |
| Search mode | Hybrid search | Settlement statement numbers and invoice numbers are exact queries, while semantic phrasings are natural-language queries, and both occur |
| Minimum relevance | Starting point 0.5 | A starting point for this trial, used to compare the number of recalled items against the default-off state; this is not an official default value |
| Citation limit | Starting point 2000 tokens | A starting point for this trial, estimated from the body length of a single settlement statement, compared with a 1000-token result |
| Result re-ranking | Enabled | A re-ranking model must already be configured; compare the ordering against the result when it is not enabled |
| Question optimisation | Enabled | Multi-turn follow-up scenarios need coreference resolution; compare the recalled content against the result when it is disabled |
Three common errors
- Treating a retrieval hit as authorisation. The input is the requester's identity and the order being asked about; the comparison is against the access scope of the knowledge base holding that order's materials; the observable result is whether the material falls inside the authorised scope. A retrieval returning a citation does not mean the requester is entitled to view that settlement statement.
- Accepting the arrival date in the draft sentence at face value. The input is the draft sentence above and material three; the comparison is the credit date field stated in each; the observable result is that March 2025 and 2025-04-05 disagree, which is judged to be a mutually exclusive pair of values for the same field.
- Writing a pending-review status as approved. The input is the authorised disbursement field, which the materials do not state; the comparison is whether the draft sentence contains that conclusion; the observable result is that the draft sentence has no support in the materials and is kept as pending verification.
How to verify the setup
- Check the draft sentence against material one and material three. The input is the two passages of original text above and the draft sentence; the comparison covers the settlement amount and the credit date fields; when the date in material three is 2025-04-05, the claim in the draft sentence that the money arrived in March 2025 is judged not to hold.
- Check downstream handling with a clearly marked paper input of
[]. The input is an empty citation array output by the knowledge base search node; the comparison is whether the downstream answering step still runs and states honestly that no material was retrieved; the observable result is that the chain is not interrupted and the answer contains no source claim. - Check the exchange-rate citation against material two. The input is the original text of material two and the part of the draft sentence that refers to the exchange rate; the comparison is between the applicable period 2025-03-01 to 2025-03-31 and the period stated in the draft sentence; when the periods agree, the rate citation holds, and when they disagree, the citation is judged to be of the wrong version.