What the documents look like
Settlement and foreign-exchange materials for cross-border platform sellers revolve around funds already collected from completed orders. Fraud and chargeback handling sits outside this scope. The working sources are the platform settlement center, bank statements, payment institution back office, and ERP orders. Day-to-day documents include platform settlement statements, exchange-rate tables, service-fee invoices, and receipt reconciliation sheets. These documents appear in Chinese, English, or the language of the platform site, and one settlement may carry fields in more than one language at the same time. The retrieval task is to locate checkable source text inside these materials so finance staff can compare order deductions, exchange-rate differences, and service-fee aggregation. Material updates follow a weekly or monthly rhythm; this page does not cover batch synchronization or real-time triggers.
Constraints at the vector indexes step
Knowledge base search runs against the whole library. A collection is a classification layer and does not affect search results. Knowledge bases linked for retrieval must share the same vector model; different models cannot be linked together. The practical step is to check the actual model and input type first, then exclude any library that does not match. One original data block can carry several indexes, and the index content is checked separately from the source text. Index content should be complete and focused, and the trade-off between retrieval coverage and index length is weighed against sample material. Image retrieval depends on whether an image description or image vector index was generated at ingestion time, so the query side alone does not decide the outcome.
Semantic retrieval suits natural-language questions and cross-language matching. Full-text retrieval suits numbers, model codes, and proper nouns. Hybrid retrieval merges multiple recall paths through RRF, and whether a rerank pass follows depends on the configured setting. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; it is a filtering threshold and does not perform permission or entity isolation. The same-vector-model requirement for linked libraries does not extend to business-registration sites or other external systems.
A hypothetical drill sample is used below. The source lines for comparison are: "Settlement statement CN-2024-07 records order A1001 service fee USD 12.00, exchange rate 7.10, settlement date 31 July 2024." and "The July 2024 exchange-rate table lists a USD/CNY central parity of 7.10 for 31 July." The draft sentence to be judged is: "The settlement rate for order A1001 matches the rate recorded in the exchange-rate table for 31 July 2024." The handover request is: "Please confirm whether service-fee aggregation for this order has been completed."
Configuration and implementation checks
The table lists documented settings or implementation checks tied to the drill sample above.
| Configuration or check | Value or state in this sample | Note |
|---|---|---|
| Linked knowledge base | Settlement materials library, holding the Chinese settlement statement and the English exchange-rate table | Confirm the library uses one vector model |
| Search mode | Hybrid retrieval | Semantic recall plus full-text recall, merged by RRF |
| Result reranking | Enabled | Requires a rerank model, enabled per configuration |
| Reference cap | Hypothetical trial start of 3000 tokens | Chosen to compare cited source text across the same question set against a smaller value |
| Minimum relevance | Hypothetical trial start of 0.35 | Applies only under semantic retrieval or reranking; compare against a higher value to observe the filter range |
| Question optimization | Off | The sample question has a complete subject and predicate; off avoids an extra model call |
Three common errors
- Linking libraries that use different vector models. The input is two knowledge bases with different vector models; the comparison is the actual model identifier of each library; the observable result is that an identifier mismatch means the link does not meet the requirement, so the mismatched library is excluded and the actual observation is recorded without inferring runtime behavior.
- Treating minimum relevance as a permission or entity filter. The input is a library holding settlement statements for several entities plus a low relevance threshold; the comparison is whether returned references still contain records for other entities; the observable result is that the threshold changes only the relevance filter range and does not exclude unauthorized records, which require a separate system implementation.
- Treating image retrieval as a query-side capability alone. The input is a settlement voucher containing an image plus a pure-image query; the comparison is whether an image description or image vector index was generated at ingestion; the observable result is that when no such index exists at ingestion, the image cannot be retrieved, so the ingestion index state is checked first.
How to verify the setup
The steps below use the hypothetical drill sample and do not claim that any test was performed.
- Input the draft sentence together with the two source lines above; compare the subject (order A1001), the period (31 July 2024), and the field (exchange rate 7.10); the observable result is that when all three agree the sentence is supported by the source text, and when any one is missing the item is recorded as pending verification.
- Input the handover request together with the sample source lines; compare whether the source text states the service-fee aggregation status; the observable result is that the source lines state only the service fee amount and the exchange rate and do not state whether aggregation is complete, so the request remains unsupported and a pending review does not demonstrate a lack of operational authority.
- Input a pure-image query against the sample library; compare whether an image description or image vector index was generated at ingestion; the observable result is that when no such index exists at ingestion the image cannot be retrieved, so the ingestion index state is checked first; an empty result can be checked against the downstream design using the explicit paper input
[], without asserting that a fictional rule search always returns[].
The public technical facts confirmed for this feature are: linked knowledge bases must share one vector model, so the actual model and input type are checked first; one original data block can carry several indexes, and index content is checked separately from source text; index content should be complete and focused, with coverage and length weighed against sample material; image retrieval also depends on whether an image description or image vector index was generated at ingestion, so the query side alone does not decide the outcome; after a model or index change, the same question set is used to compare cited source text, and a vector score alone does not promise an outcome.
This page covers only cross-border platform settlement and foreign exchange, cross-border materials and multilingual retrieval, and vector models and indexes. Which materials, operations, and requests sit at what authorization state during an operations handover cannot be treated as authorized execution through retrieval or a form.