What the documents look like
This category covers foreign-exchange receipt and settlement together with trade directory records. The working materials are receipt vouchers for incoming foreign exchange, trade contracts, invoices, and foreign-exchange business registration records. The people handling them are trade finance staff, bank relationship managers, and foreign-exchange specialists. The purpose differs from letter-of-credit document examination: the check concerns the authenticity of the foreign-exchange funds and the underlying trade background, without reviewing a full set of letter-of-credit documents. Records appear per transaction and per month, and a single receipt of funds corresponds to one set of voucher, contract, and invoice. These facts define the input boundary for copy drafting. Quotable content is limited to the counterparties, amounts, goods descriptions, and registration status already stated in those materials. Anything the materials do not state stays unknown.
A hypothetical paper exercise sample is used throughout this page. Material text one reads: "Exporter: a machinery company in East China; importer: an overseas trading company; goods: 200 small water pumps; contract number HX-2405; invoice number FP-2405-18; the receipt voucher shows USD 86,000 credited on 20 May 2024." Material text two reads: "The foreign-exchange business registration record states that this receipt corresponds to export income under trade in goods, and the registration status is registered." The draft sentence to be judged is: "The company is a large exporter whose products sell worldwide, and settlement was completed on the day the funds arrived." The review task is to decide whether the draft sentence is supported by the two material texts and to note which fields would need to be supplied.
Constraints at the multi-turn prompts step
Multi-turn dialogue is used to fill in the material scope step by step, confirm where each quoted item comes from, and stop when evidence is insufficient. The prompt needs to state the material scope, the answering task, and how to behave when evidence is missing, so that the model does not write in business facts that were never supplied. The history setting determines which earlier turns are carried into context, and a new session differs from a continued session in what is sent, so the two must be checked separately. Question optimization is checked for reference resolution and question expansion only when continuous knowledge-base dialogue is required; short questions do not necessarily call for it. Model and length are checked against the context and reply settings actually supported. Reply format follows what the downstream step needs; when json_schema is selected, the corresponding format requirement is filled in and the returned structure is verified.
The sample above fixes the evidence boundary. The exporter name, importer name, goods, quantity, contract number, invoice number, amount, currency, and value date appear in the materials. The registration record states the trade category and the registration status. The draft sentence adds company scale, worldwide sales reach, and same-day settlement. None of those three items appears in either material text. The review therefore compares each business fact in the draft sentence against the material texts for the same counterparty, the same period, and the same field. Fields that are absent remain unsupported rather than contradicted.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| AI model | A model shown in the configuration dialog as supporting a long context and tool calling | Compare the displayed maximum context and tool-calling capability against the combined voucher, contract, and invoice material |
| Memory turns | Keep a small number of earlier turns | Compare against the earliest turn needed for the follow-up questions; a larger number brings in unrelated content |
| Reply limit | Enabled, with the reply length limited | Compare against the length the draft sentence requires; too low truncates, too high increases consumption |
| Temperature | On the low side | Compare against stable customer-service and knowledge-base question-answering scenarios; divergent copy needs are assessed separately |
| Reply format | Keep the default | Compare against whether the downstream step needs a fixed structure; select json_schema when it does |
| Multimodal recognition | Enable image recognition according to the model capability | Compare against whether the materials contain screenshots or table images; document content still goes through file parsing |
These values are hypothetical starting points for this exercise, not official defaults. The configuration dialog shows only the options the current model supports. When a model does not support multimodal recognition, that option is not offered, and when it does not support thinking configuration, those options are not shown either.
Three common errors
- Writing company scale, sales reach, or settlement timing into the draft sentence when the materials do not state them. Check: take the two material texts above and the draft sentence, compare whether "large exporter", "sell worldwide", and "settlement on the day the funds arrived" appear in either material text, and mark each as unsupported when it does not appear.
- Treating a similarity score as a permission or entity filter. Check: submit one query unrelated to the sample, compare the returned records against the sample counterparty, and when they differ record the result as a retrieval outcome without concluding that access was exceeded or that records were excluded.
- Treating the display configuration returned by an interactive node as a business action already carried out. Check: submit a request that reaches an interactive node, compare whether the returned type and params contain display fields only, and record internal runtime fields such as entryNodeIds as not returned.
How to verify the setup
- Enter the hypothetical material texts above, the draft sentence, and the review task. Compare each business fact in the draft sentence against the material texts for the same counterparty, the same period, and the same field. The sentence meets the supported condition only when every fact maps to a corresponding field; otherwise record the gap. The expected observation is that company scale, worldwide sales reach, and same-day settlement are each recorded as unsupported.
- Enter an empty-result scenario using the explicit paper input []. Check whether the downstream step handles an empty result as designed. Do not claim that any rule returns [] in a live retrieval.
- Enter knowledge-base identifiers that use different embedding models. Compare whether the embedding-model identifiers of the knowledge bases match. Exclude those that do not match and record the actual observation. Do not run a cross-model retrieval to compare missed recall.
The same-model requirement for a shared retrieval association applies to the knowledge bases being linked. It does not extend to business-registration websites or other external systems. A relevance score is not a category, entity, or permission filter, and retrieval does not by itself grant authorization. Forms collect values; business verification, approval, database writes, and external synchronization each need their own downstream implementation. When the materials do not state a field, that field stays unknown, and no authorization or effective clause is invented to reach a definite answer.