What the documents look like
Cross-border platform settlement and exchange work covers the collection of funds from completed orders. It does not handle fraud chargebacks; that belongs to cross-border payment risk control. The working materials are platform settlement statements, exchange rate tables, service fee invoices and receiving reconciliation tables. Settlement data updates weekly or monthly, and the materials arrive at different times. A platform settlement statement may be issued first, with bank statements and receiving reconciliation tables following later.
In store service and internal training, staff explain to colleagues or clients which document, which field and which period a given amount, rate or fee comes from. Citation tracing requires every quoted item to return to a specific document, a specific field and a specific period. When materials are incomplete, checking can only cover the parts already obtained, and the gaps are recorded as they stand.
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. An empty array means the output path still runs; it does not mean the workflow has errored out or stopped. Citation content then passes to the answering step, where each answer sentence is checked against the cited source text.
The minimum relevance value is a filter threshold. It takes effect only under semantic retrieval or when result reranking is used, and it does not perform permission or entity isolation. The same-vector-model requirement constrains knowledge base association only; it does not extend to external systems needing the same model.
Hypothetical exercise sample: a source text titled Platform Settlement Statement for June 2025 reads "settlement period 2025-06-01 to 2025-06-30, net settlement amount CNY 12,340.00, deduction items are commission and service fee." A source text titled Receiving Reconciliation Table reads "2025-06-30 received CNY 12,340.00, payer is the platform settlement center." A draft sentence to be judged reads: "The store's June settlement has been received in full with no deductions." The source texts support only that the received amount equals the net settlement amount and that deduction items are listed. The phrase "with no deductions" does not match the source text.
Configuration and implementation checks
The values below are starting points for this exercise, used to compare against the sample source texts. They are not official defaults.
| Configuration item | Starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | One settlement statement base and one reconciliation table base | Both bases may be associated together only when their vector model identifiers match; check the identifiers one by one |
| Search mode | Hybrid retrieval | Settlement numbers and amounts are exact fields, so semantic and full-text recall run in parallel |
| Citation limit | 3000 tokens | Combined length of one settlement statement plus a reconciliation table excerpt; compare against the actual retrieval output length |
| Minimum relevance | 0.4 | Takes effect only under semantic retrieval or reranking; compare against the rerank score distribution before adjusting |
| Result reranking | Enabled | Secondary ordering when candidates are many; compare against the ordering without reranking |
| Citation handoff | Passed to the answering node | Check each answer sentence against the cited source text and confirm the fields match |
Three common errors
- Treating retrieval output as an authorization conclusion. The search returns a citation array, which shows only that material fragments were found. It does not show that the payment was approved or credited. The check input is the retrieval output array; the comparison object is the actual crediting record; the observable result is whether both point to the same amount and the same period.
- Treating minimum relevance as an entity filter. This value filters low-relevance results; it does not exclude unrelated businesses or unauthorized records. The check input is the relevance setting; the comparison object is the business entity in the returned results; the observable result is whether returned content still contains records of other entities.
- Writing a pending state as completed. If a draft sentence says "received in full with no deductions," it must be compared against the deduction item field of the settlement statement. The input is the draft sentence; the comparison object is the deduction column of the settlement statement; the observable result is whether the draft sentence matches the deduction items listed in the source text.
How to verify the setup
- Enter the settlement statement source text and the reconciliation table source text from the hypothetical sample. Compare the amount field and the period field of both. The condition is met when the amounts match and the periods match. If the draft sentence contains "with no deductions," compare it against the deduction column of the settlement statement and judge it as unsupported.
- Enter a question containing only a settlement number. Compare whether the returned citation points to the corresponding settlement statement fragment. The condition is met when the citation content contains that number. If an empty array is returned, record it as not recalled; do not judge it as an error.
- Enter an association request across two knowledge bases with different vector models. Compare the vector model identifiers of the two bases. The condition is met when the association is excluded because the identifiers do not match. Record the actual observation; do not infer runtime behavior after association.