What the documents look like
Settlement and foreign-exchange work handles the collection of funds from completed orders. It sits in a different stage from payment risk control, which deals with fraud and chargebacks. Operating reports and exception reviews draw on four materials: platform settlement statements, exchange-rate tables, service-fee invoices, and collection reconciliation sheets. These materials update weekly or monthly.
The four materials come from different systems. A platform settlement center, bank statements, a payment provider back office, and ERP orders each produce their own records. The same deduction can sit in different states across a platform settlement statement and a collection reconciliation sheet: deducted, pending settlement, or posted. Citation tracing requires stating which document, which period, and which state a cited figure comes from. A figure from one state cannot serve as evidence for another state.
The task for citation tracing is specific: in a review report, mark which document and which location each amount or variance comes from, then check that location's version and scope of application.
Constraints at the citation tracing step
When a knowledge base search node links knowledge bases, it can only select knowledge bases that share the same vector model. Knowledge bases with different vector models cannot be linked together. The handling is to check the vector model identifier and exclude those that do not match, record the actual observation, and not run cross-model retrieval to compare missed recalls. The search node outputs citations as an array, and the length can be zero. An empty array does not mean the workflow has errored or terminated. Retrieval results do not constitute authorization. Forms collect values; business verification, approval, database writes, and external synchronization each need their own downstream implementation.
Hypothetical paper exercise sample (paper materials, not real records):
Material text one, from a hypothetical platform settlement statement: "Settlement period 2025-03-01 to 2025-03-31, order number A-1001, deduction 12.00 USD, status: deducted."
Material text two, from a hypothetical collection reconciliation sheet: "2025-03-31 posted A-1001 net amount 11.40 USD, service fee 0.60 USD."
Draft sentence to be judged: "A-1001 completed all settlement in March 2025, with no outstanding amount."
Judgment points: both materials share the same subject (A-1001), the same period (March 2025), and the same fields (settlement status and amount). The settlement statement records a deduction of 12.00 USD. The reconciliation sheet records a posted net amount of 11.40 USD and a service fee of 0.60 USD. The draft sentence says "completed all settlement." The materials do not record a confirmation field for completed settlement, so this claim is not supported. The draft sentence says "no outstanding amount." The materials do not record an outstanding amount field, so this claim is also not supported. The two amounts can corroborate each other on the deduction composition, but they do not confirm "completed all settlement."
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for choice and comparison method |
|---|---|---|
| Linked knowledge bases | Select two knowledge bases with matching vector model identifiers | The document requires the same vector model for joint linking; compare each knowledge base's vector model identifier and exclude those that do not match |
| Search mode | Hybrid search | Settlement statements contain exact order numbers suited to full-text recall; variance descriptions suit semantic recall; compare with semantic search alone |
| Citation limit | Hypothetical starting point 3000 tokens | The document measures citation limit in tokens; compare citation count and truncation against 1500 and 6000 |
| Minimum relevance | Hypothetical starting point 0.4 | The document states this value takes effect only in semantic search or with result reranking; compare filtered count against 0.2 and 0.6 |
| Result reranking | Enable when a rerank model is available | The document states reranking is enabled by setting; compare ordering with and without it |
| Citation content output | Accept the array, allow length zero | The document states the empty-array path still runs; check the downstream design for handling empty arrays |
Three common errors
- Treating a retrieval hit as authorization. A retrieval returning a settlement statement citation does not mean the funds have been approved for payment or posted. Error symptom: in a review, using "retrieved" in place of "approved." Check: input the hypothetical sample above, compare whether the draft sentence writes the existence of a citation as a completed business state, and observe whether completion wording appears that the materials do not record.
- Forcing knowledge bases with different vector models into a joint link. Error symptom: selecting knowledge bases with mismatched vector model identifiers and expecting joint retrieval. Check: verify each knowledge base's vector model identifier, exclude those that do not match, record the actual observation, and do not run cross-model retrieval to compare missed recalls.
- Treating an unrecorded field as confirmed. Error symptom: the materials only have deduction and posting records, but the draft sentence says "completed all settlement" and "no outstanding amount." Check: compare the material text against the draft sentence field by field, and mark fields the materials do not record as unsupported.
How to verify the setup
- Input the hypothetical sample material text one and text two, compare "completed all settlement" in the draft sentence against the settlement status field recorded in the materials, and observe whether the claim is supported by the materials. When the materials do not record a completion field, judge it as unsupported.
- Input "no outstanding amount" from the draft sentence, compare whether the materials record an outstanding amount field, and observe whether the claim is supported by the materials. When the materials do not record that field, judge it as unsupported.
- Input two hypothetical knowledge bases with mismatched vector model identifiers, compare the vector model identifiers, and observe that the handling is to exclude those that do not match and record the observation, without running cross-model retrieval to compare missed recalls. At the same time, use an explicit paper input of an empty array to check the downstream design for handling empty citations, without claiming that a fictional rule's retrieval necessarily returns an empty array.