What the documents look like
Settlement and foreign-exchange work covers the collection of funds for orders that have already been completed. It does not handle fraud or chargeback disputes. The materials a draft draws on are therefore platform settlement statements, exchange-rate tables, service-fee invoices and receiving reconciliation sheets. These materials update weekly or monthly.
For copy drafting, the practical consequence is that any figure, rate, exchange rate or payment-term wording that appears in a draft needs to trace back to the exact wording and version of one specific material. A settlement statement shows a fee amount for one period. An exchange-rate table shows a value for one date. Neither of these is the same as a rate commitment or a standing conversion standard. Tracing a citation means confirming the material's subject, period and field first, then judging whether the proposed sentence is supported by that field.
Constraints at the citation tracing step
The knowledge-base search node outputs cited content as an array, and the length may be zero. An empty array means only that this search returned no citations; the workflow still continues, so a downstream step has to handle that branch. A search can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together, so the identifiers must be checked and non-matching ones excluded, with the actual observation recorded rather than any inference about what association would produce.
The minimum relevance value is a filter threshold. It takes effect only under semantic search or when result reranking is used, and it does not perform permission or entity isolation. Retrieval itself is not authorization. Hybrid search merges multiple result paths through RRF, and reranking requires a model and must be enabled according to the configuration. Image retrieval requires corresponding image input and model capability. These paths do not mean every request executes all of them.
Hypothetical exercise sample: Material text one is "Platform settlement statement for March 2025 (number HYP-2025-03-A) states: order deductions for this period total 8,600 yuan, of which payment service fee is 1,200 yuan, currency US dollars." Material text two is "Receiving reconciliation sheet for the same month (number HYP-2025-03-B) states: amounts received from 1 March to 31 March total 7,400 US dollars, status posted." The sentence to be judged is "The store's payment service fee is fixed long-term at 1,200 US dollars per period." The judgment needs the same subject, the same period and the same field. The two materials support only the service-fee amount for the March 2025 period. They do not state that it is fixed long-term, so the proposed sentence is unsupported and the evidence gap remains.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Select only the settlement and exchange knowledge bases | Check whether the vector-model identifiers of the selected bases match; exclude non-matching ones |
| Search mode | Hybrid search | Compare with semantic search alone on hits for numbers and amount fields |
| Minimum relevance | Hypothetical starting point 0.5 | Compare with the displayed value; observe whether low-relevance passages are filtered |
| Citation limit | Hypothetical starting point 4,000 tokens | Compare with the length of the material passages the draft needs; observe whether citations are truncated |
| Result reranking | Enable according to the actual configuration | Compare candidate ordering with reranking off; do not enable when no model is configured |
| Citation handoff | Check whether the output array is passed to the answering step | Compare with the empty-array branch; confirm the downstream step has a handling design |
The values above are hypothetical starting points for this exercise, not official defaults.
Three common errors
- Writing a service-fee amount from one settlement period as a fixed long-term rate, without checking whether the same field states an applicable period. Check: input material text one and the proposed sentence, compare whether "this period" and "fixed long-term" belong to the same field, and treat the condition as met only when the material states long-term applicability.
- Treating a single day's value in an exchange-rate table as the conversion standard for a whole month. Check: input the exchange-rate table's date and the proposed sentence's period, compare whether the two period ranges match, and record the result as unsupported when the periods differ.
- Writing a pending-review status as signed or already effective. Check: input the status field in the material and the wording of the proposed sentence, compare whether the status values are mutually exclusive, and note that pending review does not show an absence of operating authority and does not equal contract effectiveness.
How to verify the setup
- Input material text one, material text two and the proposed sentence from the hypothetical sample. Compare subject, period and field for consistency. The sentence passes only when the materials support it; when the materials do not state it, record the gap.
- Input an explicit paper-based empty array and check whether the downstream step has a handling design. Observe whether the flow continues. Do not claim that a fictional rule search will necessarily return an empty array.
- Input two knowledge bases whose vector-model identifiers differ, check the identifiers and exclude the non-matching one, and record the actual observation. Do not run a cross-model retrieval to compare missed recalls.