What the documents look like
Cross-border platform settlement and foreign-exchange records center on platform settlement statements, exchange-rate tables, service-fee invoices, and collection reconciliation sheets. Sources include platform settlement centers, bank statements, payment-institution back offices, and ERP orders. These materials update weekly or monthly, and field definitions are set separately by each platform and each bank.
Contract and transaction material comparison requires aligning the contracted rate, settlement cycle, and currency against the actual deductions and credited amounts shown in the transaction materials. The same order amount can appear in a settlement statement, a reconciliation sheet, and an invoice. Retrieval therefore needs constraints on three dimensions at once: entity, period, and field. When any one of these is missing, a comparable conclusion cannot be formed.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. Search runs against the whole knowledge base; a collection is only a way to organize data and does not affect search results.
Semantic search relies on vector distance. Full-text search suits identifiers, model numbers, and proper nouns. Hybrid search combines semantic and full-text recall through RRF. Rerank requires a model and is enabled according to settings. The reference limit is measured in tokens. The minimum relevance value takes effect only under semantic search or when result reranking is used. The cited content output is an array whose length can be 0, and the output path still completes when nothing is retrieved. Retrieval and similarity scores do not perform permission checks or entity isolation, and they do not determine whether a contract has taken effect or whether authorization has been completed.
Hypothetical exercise sample: the source text reads, "Party A, XX Supply Chain (example entity), and Party B, YY Logistics (example entity), signed a Cross-border Logistics Service Agreement, agreeing that the service fee rate for the March 2025 settlement period is 18 yuan per kilogram, and the settlement currency is US dollars." The draft sentence under review reads, "The service fee under this agreement for March 2025 is settled at 18 yuan per kilogram in RMB." The comparison checks the currency field: the source states US dollars, the draft states RMB, and mutually exclusive values in the same field constitute a contradiction. If the draft instead reads, "The service fee is settled at 18 yuan per kilogram, currency to be verified," the currency is unsupported and the item is an evidence gap.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Settlement statement base, exchange-rate table base, invoice base | The three bases use the same vector model; verify the model identifier first and exclude any that do not match |
| Search mode | Hybrid search | Rate identifiers and natural-language clauses coexist; compare the difference between semantic and full-text recall paths |
| Reference limit | Hypothetical starting point of 2000 tokens | Settlement statement segments are long; compare content completeness against a 500-token starting point |
| Minimum relevance | Hypothetical starting point of 0.3 | Label relevant and irrelevant samples first, then compare false recalls and misses against a 0.5 starting point |
| Question optimization | Enabled | The entity is easily dropped in follow-up turns; compare the entity in the original question and the optimized question |
| Result reranking | Enabled | When candidates are numerous, compare ordering before and after reranking and confirm the model is configured |
Three common errors
- Treating a retrieval hit as authorization confirmation. Input "XX Supply Chain March 2025 settlement statement"; retrieval returns cited content, but retrieval does not determine whether that entity may view the record or whether the contract has taken effect. Check: compare the retrieval result against the authorization record, and record the item as pending verification when the authorization record is missing.
- Treating a missing currency field as a contradiction. The draft reads "settled at 18 yuan per kilogram"; the source does not state a currency, so the item is unsupported. Check: compare the draft field against the same field in the source, and mark it as an evidence gap when the source has no such field.
- Treating a pending-review status as signed. The material shows "agreement pending review"; the draft states "agreement signed." Check: compare the status field; pending review and signed are different facts and neither may be inferred from the other.
How to verify the setup
- Input the hypothetical source text and the draft sentence "The service fee under this agreement for March 2025 is settled at 18 yuan per kilogram in RMB," and compare the currency field. The source states US dollars and the draft states RMB; mutually exclusive values in the same field should be identified as a contradiction.
- Input the draft sentence "The service fee is settled at 18 yuan per kilogram, currency to be verified," and compare the currency field. The source does not state a currency, so the output should be unsupported and must not be judged as a contradiction or a negation.
- Input an explicit paper empty result [] and check whether the downstream design handles a reference length of 0, without claiming that a fictional rule search necessarily returns []. For cross-model bases, verify the identifier first and exclude non-matching bases; do not run a cross-model search to compare missed recalls.