What the documents look like
Cross-border platform settlement and foreign exchange work concerns the collection of funds from orders that have already been completed. It does not handle fraud chargebacks, which belong to payment risk control. The working materials are platform settlement statements, exchange rate tables, service fee invoices and receipt reconciliation tables, updated weekly or monthly. The procurement and supplier records task adds purchase orders, supplier quotations, goods receipt records and payment applications to that base. Retrieval requests in this task typically turn on supplier name, purchase batch, settlement currency and reconciliation period.
These materials span long periods, carry many fields and cross several parties. When a retrieval result mixes in fragments from another supplier or another period, the accuracy of procurement and payment reconciliation suffers directly. This page therefore addresses only the configuration and checking of the knowledge base retrieval and recall step. It does not construct automatic reconciliation or automatic payment functions that the supplied documents do not support.
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. Implementation should first check the vector model identifier, exclude those that do not match, and record the actual observation. Cross-model retrieval should not be run to compare recall differences.
Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. RRF is a fusion process and has no separate on/off switch. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used, and it is a filter threshold. It does not perform permission or entity isolation. Retrieval itself is not authorization. Whether a supplier may be paid, and whether a contract has taken effect, still require downstream business implementation and human checking.
The citation limit is measured in tokens and controls the total content cited in one search. It is unrelated to a thinking budget. Question optimization performs coreference resolution and question expansion, and it can change the subject of a question, so the original question and the optimized question need to be compared. Image retrieval and reranking in the documented flow are optional paths. They require corresponding image input, model capability and enabling under the stated settings. Their presence in the flow does not mean every request executes all of them. The empty-reference output is an array whose length can be 0. Downstream design can be checked against a clearly labelled paper input of []. No claim is made that a fictional rule search necessarily returns an empty array.
Configuration and implementation checks
The table below is built around the hypothetical drill sample on this page and lists documented settings or implementation checks relevant to this feature and the procurement task. The numeric values are starting points for this hypothetical trial only, with the reason for choosing them and the value they are compared against. They are not official default values.
| Configuration item | Hypothetical value or check action in this trial | Reason for selection and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the supplier records base and the procurement records base whose vector model identifiers match | Check identifiers first and exclude non-matching ones; do not extend the same-model requirement to external systems such as business registration sites |
| Search mode | For a request containing a supplier name and purchase batch, trial semantic, full-text and hybrid modes | Compare whether returned fragments under each mode cover the same supplier and the same period; do not presume one mode suits all materials |
| Citation limit | Hypothetical starting point of 3000 tokens for this trial | The sample contains two long materials, a settlement statement and a quotation; compare against a 1500-token starting point to see whether content is truncated |
| Minimum relevance | Hypothetical starting point of 0.5 for this trial | First label relevant and irrelevant fragments in the sample, then compare remaining fragments after filtering at 0.3 and 0.7 |
| Question optimization | Enable for follow-up questions containing references, and compare the question subject before and after optimization item by item | Check whether the subject has been changed incorrectly; when the question is already clear, it may be turned off according to observed effect |
| Result reranking | Trial when a rerank model is available and enabled under the settings | Compare ranking differences between rerank scores and vector scores; when not enabled, record that this path was not executed |
Three common errors
- Treating minimum relevance as a permission or supplier isolation mechanism. This value is a filter threshold and takes effect only under semantic retrieval or result reranking. It cannot exclude unauthorized records. Whether a retrieval result may be used for payment still requires downstream business checking.
- Describing RRF as an independent switch that can be turned on, or assuming hybrid retrieval always executes image retrieval and reranking. RRF is the fusion process for hybrid results. Image retrieval requires image input and model capability, and reranking requires a model and enabling under the settings.
- Inferring missed recall or redundancy from associating knowledge bases across vector models. The documents state only that association requires the same vector model. Implementation should check identifiers, exclude non-matching ones and record actual observations, without fabricating run results.
How to verify the setup
The verification below uses the same hypothetical drill sample. Sample material text (hypothetical drill sample, not a real record): "Supplier A submitted a procurement quotation on 10 March 2025, valid until 30 April 2025; the corresponding platform settlement statement period is March 2025, and the settlement currency is US dollars." Draft sentence to be judged: "Supplier A's quotation remains valid in May 2025 and can be used directly for this payment application." The request is pending review and has not received payment authorization.
- Input the sample material text above and the draft sentence, then compare whether the validity period stated in the draft sentence and the 30 April 2025 date recorded in the source are the same field with the same value. The draft sentence is supported only if the source explicitly records that it remains valid in May. At present this should be recorded as unsupported.
- Input the follow-up question containing a reference, "What is its settlement currency", and compare whether the question subject still points to Supplier A before and after question optimization. The condition that the subject was not changed incorrectly is met only if the optimized question retains Supplier A and the March 2025 period.
- Input the sample material text above and compare whether the returned fragments cover both the March 2025 settlement statement for Supplier A and the quotation. The content completeness condition is met only if the cited content contains those fields and is not truncated by the citation limit. If the returned length is 0, check the downstream design against the paper input
[], and do not infer that the retrieval rule necessarily returns an empty array.