What the documents look like
Cross-border platform settlement and foreign exchange materials centre on funds collection for orders that have already been completed. The main sources are the platform settlement centre, bank statements, payment institution back office, and ERP orders. Day-to-day documents are platform settlement statements, exchange rate tables, service fee invoices, and receipt reconciliation tables, with updates falling mostly on weekly and monthly cycles. In store service and internal training, these materials support two retrieval families: field-level questions about settlement statements and reconciliation tables, and collection-basis questions about exchange rate tables and service fee invoices. The people handling them are cross-border seller finance staff, tax accountants, and platform operations staff. Errors surface in profit calculation, tax filing, and cash planning. Retrieval here has to accept a mix of identifiers, currencies, periods, amount fields, and natural-language phrasing, so semantic search and full-text search each have a place.
Two or three of these facts shape the retrieval task directly. The first is that the smallest search unit is the library, and collections inside it only classify data rather than changing search behaviour. The second is that the material mixes exact identifiers and amounts with open questions about treatment, which is why search mode selection matters. The third is that the update cadence is weekly and monthly, so a training answer built on a retrieved statement is only as current as the statement version in the library.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Search modes are compared against test questions across semantic, full-text, and hybrid retrieval; no single mode is assumed to suit all materials. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion and calls a model an extra time. Image retrieval requires image input and the corresponding model capability. Reranking requires a model and must be enabled by configuration. RRF is the fusion process for mixed results, not a standalone switch. Retrieval itself does not perform authorization checks, and a similarity score is not an entity or permission filter.
Hypothetical drill sample. A store internal training pack contains a statement whose text reads: "Settlement entity: Xingqiao Cross-border Store; settlement period: 1 March 2025 to 31 March 2025; currency: USD; net settlement amount: 12,480.00." A receipt reconciliation table in the same pack reads: "Receiving entity: Xingqiao Cross-border Store; reconciliation period: March 2025; credited currency: USD; credited amount: 12,480.00." A draft sentence awaiting review reads: "Xingqiao Cross-border Store's March 2025 net settlement amount matches the credited amount, so no difference in that month's funds collection can be confirmed." The entity, period, currency, and amount fields in the draft all have corresponding values in the source text, but the "no difference" conclusion requires checking whether both materials cover the same complete period and whether any fee deduction items are unlisted.
Configuration and implementation checks
The configuration below relates to the sample and lists only settings or implementation checks that the official documents describe for this feature and the store training task.
| Configuration item | Hypothetical starting point for this drill | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Statement library and reconciliation library; check vector model identifiers first | Both libraries may be associated together only when their vector model identifiers match; exclude any that do not |
| Search mode | Run one pass with semantic retrieval and one with full-text retrieval | Identifiers and currency fields suit full-text; period and treatment questions suit semantic |
| Citation limit | Start at a token amount that covers the complete fields of one settlement statement | Compare against the actual length of a single statement; raise the limit if it is insufficient |
| Minimum relevance | Start at 0.5, effective only under semantic retrieval or reranking | Adjust after comparing against labelled relevant and irrelevant samples |
| Question optimization | Enable for short follow-up questions; disable for complete field questions | Compare the original question with the optimized question and check whether the entity was altered |
| Result reranking | Enable when candidates are numerous, according to configuration | Compare ordering before and after reranking and confirm the text question is complete |
Three common errors
- Treating a retrieval hit as authorization approval. The draft sentence in the sample asks for confirmation that funds collection shows no difference; retrieval can only return the statement and reconciliation text, while business confirmation and approval need a separate system implementation. The observable error is that once cited content appears, the reviewer declares it confirmed. The input is the retrieval result array, the comparison object is the authorization status field, and the observable result is that the authorization field remains empty.
- Extending the same-vector-model requirement to external systems. The association requirement for knowledge bases constrains the knowledge base search node only, not a business registration site or a bank system. The observable error is requiring external systems to use the same model. The input is the external system identifier, the comparison object is the node association condition, and the observable result is that the external system identifier falls outside that condition.
- Treating an unstated field as negative evidence. The reconciliation table in the sample does not list service fee deductions, and the draft sentence concludes there is no difference on that basis. The input is the source text of both materials, the comparison object is the value of the same field, and the observable result is that the deduction field appears in neither material, so the finding must be recorded as unsupported or pending verification.
How to verify the setup
- Input the source text of the March 2025 platform settlement statement and the March 2025 receipt reconciliation table from the sample, and compare the entity, period, currency, and amount fields. When all four fields take the same values in both materials, the retrieved citation content should contain both source texts. If any field is missing, record it as an evidence gap and do not judge the values consistent.
- Input the draft sentence "Xingqiao Cross-border Store's March 2025 net settlement amount matches the credited amount, so no difference in that month's funds collection can be confirmed," and compare the draft's conclusion against the actual coverage of the two materials. When neither material lists service fee deductions, the conclusion is unsupported. Only when a material lists deductions that can be matched to the net amount does the check move to the next step.
- Input an empty citation array [] and check the downstream design. The node's citation output length can be 0, and the empty-result path still completes. Observe whether the downstream handles the empty result explicitly; do not claim that a particular rule always returns []. For knowledge bases with different vector models, check the identifiers first and exclude those that do not match, rather than running a cross-model retrieval to compare missed recalls.