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 working materials are platform settlement statements, bank statements, payment-provider back-office records, and ERP orders. These four sources form the reconciliation base, and each one carries its own version and period.
Platform settlement statements frequently contain multilingual fields and localized fee names. A fee item may appear under a translated label that differs from the label used in the ERP order or the service-fee invoice. The update rhythm is weekly or monthly, which means a citation must be tied to a specific document version. The same order can show different amounts across periods when the exchange-rate table is revised. The field correspondence between service-fee invoices and receipt reconciliation sheets is the first object that needs confirmation during tracing.
For citation tracing, this matters because a retrieved passage is only useful when the reader can identify which statement version, which fee label, and which period it came from. A fluent answer that cites a fee name without the matching version cannot be checked against the original.
Constraints at the citation tracing step
The knowledge-base search node can associate one or more knowledge bases that share the same vector model. Citation content is output as an array, and its length can be zero. A zero-length array still passes through the output path. The retrieval result must be handed to the answering stage, where each answer is checked line by line against the cited original text. A smooth answer does not replace source evidence.
Semantic retrieval and relevance filtering do not carry permission or entity-isolation duties. The same-vector-model requirement for associated knowledge bases does not extend to external business-registration systems or any other outside system. Mixed retrieval combines vector and full-text recall through RRF, and RRF is a fusion process, not a switch. Re-ranking requires a re-rank model and must be enabled in the settings. Image retrieval requires an image input and a model with the corresponding capability. The optional paths in the overall retrieval flow are not all executed on every request.
Hypothetical paper exercise: a cross-border seller's finance staff member plans to check the "FBA storage fee" item in the March 2025 platform settlement statement. The source text reads: "Settlement statement number ST-202503-A, subject is a Shenzhen trading company, fee item FBA storage fee, amount converted at the March 2025 exchange-rate table, status settled." A draft sentence reads: "This FBA storage fee has been deducted from the March 2025 settlement statement and can be included in the current month's cost." The question to judge: does "has been deducted" in the draft sentence belong to the same field as "settled" in the source text, as a mutually exclusive value?
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Select the base containing settlement statements and exchange-rate tables | Check whether the selected bases share the same vector-model identifier |
| Search mode | Hybrid retrieval | Fee names contain proper nouns, so semantic and full-text recall are used together |
| Citation limit | Measured in tokens, starting point 4000 | Compare against the text length of a single settlement statement, then adjust |
| Minimum relevance | Starting point 0.5 | Compare against the 0-1 re-rank score range and observe the filtering effect |
| Result re-rank | Enabled | Secondary ordering when candidates are numerous; compare against results without re-rank |
| Question optimization | Not enabled for now | Single-turn explicit question; compare against multi-turn follow-up scenarios |
The citation limit is expressed in tokens because chunk lengths vary widely across a mixed knowledge base, and a fixed top-k count produces unstable results. The minimum relevance value applies only under semantic retrieval or when result re-ranking is used. It is a filtering threshold, not a final ordering rule.
Three common errors
- Treating retrieval output as an authorized action. The citation array returns settlement-statement content, which does not mean the fee has received payment authorization or can be booked. Check the input as the citation array, compare against the authorization-status field, and the observable result is that the authorization field remains unverified.
- Treating a similarity score as a permission filter. Minimum relevance filters low-relevance results only. It does not exclude unrelated businesses or unauthorized records. Check the input as the relevance threshold, compare against the business ownership of returned records, and the observable result is that ownership still requires manual verification.
- Associating knowledge bases across different vector models. Only bases with the same vector model can be associated together. Check the input as the model identifier of each selected base, compare against the vector-model field, and the observable result is that bases with mismatched identifiers are excluded. Do not then run a cross-model retrieval to compare missed recalls.
How to verify the setup
- Citation tracing check. Input the hypothetical source text and the draft sentence given above. Compare whether "settled" and "has been deducted" belong to the same field as mutually exclusive values. The source text states only "settled" and does not state a deduction action. The draft sentence's "has been deducted" is unsupported and should be recorded as pending verification.
- Empty-result continuation check. Input an explicit paper array []. Compare whether the downstream answering node still passes through the output path. The node documentation states that a citation array of length zero still passes through the path. The observable result is that the downstream stage does not stop because of the empty array. This check uses a paper input and does not claim that a real query returns an empty array.
- Document-version check. Input the current version and a historical version of the settlement statement. Compare whether the citation source is bound to a valid version. If the historical version and the current version differ in amount basis, the observable result is that the citation must be reviewed against the valid version, and the version difference is recorded as pending verification.
Across all three checks, the same subject, the same period, and the same field are required before a contradiction can be called. Where a field is not stated, or only related-stage evidence exists, the result is recorded as unsupported or pending verification. A pending review does not show that an operation permission is missing.