What the documents look like
Operating reviews in cross-border platform settlement and foreign exchange draw on four recurring document families: platform settlement statements, exchange rate tables, service fee invoices, and collection reconciliation tables. These materials originate across the platform settlement center, bank statements, payment institution back offices, and ERP orders. Settlement work concerns the collection of funds for orders that have already been transacted; it sits apart from fraud and chargeback handling. Updates follow a weekly and monthly rhythm, and the fee deduction, exchange rate difference, and service fee collection for a single order are frequently spread across several documents. An operating review therefore has to align subject, period, and field across materials before any conclusion is drawn.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. During retrieval, vectors are recalled first, then the original data is located in the MongoDB dataset.datas table by vector ID. When vectors map to the same group of original data, the results are merged and the highest vector score is kept. Multi-vector mapping lets one group of data correspond to several vectors, so recalling any one of them brings out that group. Semantic retrieval, full-text retrieval, and hybrid retrieval are compared against test questions; no single mode is assumed to suit every material. Hybrid retrieval merges multiple result paths through RRF, and reranking requires a model and must be enabled in the settings. The reference limit is measured in tokens. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion; the original question and the optimized question are compared to check whether the subject was altered. Image retrieval requires a corresponding image input and model capability, and the optional paths shown in the overall document flow diagram do not mean every request executes all of them. Retrieval and similarity scores do not perform permission or entity isolation, and the same-vector-model requirement for associated knowledge bases does not extend to external systems.
Hypothetical paper exercise sample
Material text one for checking: "The March 2025 platform settlement statement records that the total order fee deduction for Store A was 12,480 yuan, the settlement currency was US dollars, and the settlement period was 2025-03-01 to 2025-03-31."
Material text two for checking: "The March 2025 collection reconciliation table records that the amount received for Store A in the same period converted to 89,760 yuan at the 2025-03-15 mid-rate of 7.19."
Draft sentence to be assessed: "Store A completed all settlement in March 2025, with no service fee difference."
Assessment note: material text one and material text two give fee deduction and receipt fields for the same subject and the same period. Neither states a conclusion about a service fee difference, and neither states an authorization or confirmation status for "all settlement completed." The draft sentence exceeds what the materials contain and is unsupported.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relationship to the sample |
|---|---|---|
| Associated knowledge bases | Associate one or more knowledge bases that use the same vector model | After the settlement statement, rate table, invoice, and reconciliation table are stored, check the identifiers against the same vector model |
| Search mode | Compare semantic, full-text, and hybrid retrieval against test questions | Exact fields such as order number and settlement period suit full-text or hybrid retrieval |
| Reference limit | Measured in tokens | The sample contains a fee deduction passage and a receipt passage; check whether both are referenced together |
| Minimum relevance | Takes effect only in semantic retrieval or when result reranking is used | Label relevant and irrelevant samples before adjusting the threshold |
| Question optimization | Used for coreference resolution and question expansion | Check whether references such as "the same period" and "that store" are completed to the correct subject |
| Result reranking | Requires a model and must be enabled in the settings | When candidates are numerous, text results are reordered and the score is used for filtering |
Three common errors
- Treating a retrieval hit as business authorization. For the input "Store A March 2025 service fee difference," retrieval returns a reconciliation table fragment. What is observed is only the referenced content. Whether a service fee difference exists and whether settlement is complete still require manual checking of the material fields. The retrieval result does not constitute an approval or confirmation.
- Treating minimum relevance as entity isolation. For a query containing "Store A," raising the minimum relevance reduces irrelevant fragments. What is observed is only the relevance filtering effect. Whether records for other stores or unauthorized records are excluded is outside the responsibility of that parameter.
- Treating RRF as a separate switch that can be turned on. After hybrid retrieval, results are merged through RRF. What is observed is the ordering after multiple candidate paths are combined. RRF is a fusion process, and the configuration interface does not provide a separate enable item for it.
How to verify the setup
- Enter hypothetical sample material text one and material text two, and query "Store A March 2025 fee deduction and receipt." Compare whether the referenced content contains both the fee deduction total of 12,480 yuan and the receipt amount converted to 89,760 yuan. The condition is met when both passages appear in the referenced content and the subject and period match the original texts. The condition is not met when only one passage appears or the subject is rewritten to another store.
- Enter the draft sentence "Store A completed all settlement in March 2025, with no service fee difference." Compare whether the draft sentence exceeds the scope of the fee deduction field in material text one and the receipt field in material text two. The condition is met when the draft sentence is marked as unsupported and requiring additional service fee difference materials. The condition is not met when the draft sentence is output directly as a confirmed conclusion.
- Enter "Store A March 2025 exchange rate conversion," and compare whether the rate value in the referenced content matches the 2025-03-15 mid-rate of 7.19 in material text two. The condition is met when 7.19 appears in the referenced content with the conversion date indicated. When another rate value appears or no date is indicated, record it as pending verification. Empty result branch: if the returned referenced content length is 0, follow the node's existing behavior that the output array may be empty, and pass the empty array to the downstream design for checking, without inferring that the retrieval rule necessarily returns empty.