What the documents look like
Order fulfilment and after-sales support in foreign-exchange receipt, settlement and trade directory work draws on four document families: foreign-exchange receipt vouchers, trade contracts, invoices, and foreign-exchange business registration materials. Compared with letter-of-credit document examination, this category focuses on the authenticity of foreign-exchange funds and the underlying trade background, and does not examine a full set of letter-of-credit documents. Typical practitioners are foreign-trade finance staff, bank relationship managers, and foreign-exchange specialists. Materials update per transaction and per month.
These facts shape what a multi-turn prompt may do. The assistant answers only within the supplied materials, states gaps when evidence is missing, and does not write business facts the materials do not contain. When a receipt voucher and a contract are both present, the prompt still has to treat the question of trade-background authenticity as something to be checked, not something already settled.
Constraints at the multi-turn prompts step
Multi-turn dialogue must check the context actually carried in, and distinguish a new session from a continued one. When continuing, the interface retrieves the session from the database by chatId, uses the last element of the messages array as the user question, and ignores the remaining messages. When chatId is not passed, context is built entirely from the supplied messages. The prompt must state the material scope, the answering task, and how to handle missing evidence. Question optimisation, where continuous knowledge-base dialogue is needed, requires checking reference resolution and question expansion. Model and length settings require checking the model's supported context and reply settings. Reply format follows downstream consumption; when json_schema is selected, the corresponding format requirement is filled in and the returned structure is verified.
Hypothetical exercise sample: a foreign-trade enterprise finance staff member submits a scanned foreign-exchange receipt voucher and a trade contract, and asks the assistant to answer a question about the trade background behind one receipt. The material text available for checking reads: "The foreign-exchange receipt voucher states payer Company A, payee Company B, amount USD 50,000, date 10 March 2025. The trade contract states seller Company B, buyer Company A, goods are one batch of mechanical parts, contract amount USD 50,000, signing date 20 February 2025." The draft sentence to be assessed is: "The trade background of this receipt has been confirmed as authentic, and settlement can be processed." This sentence turns a matter still to be verified into a completed conclusion, and is a draft that needs correction.
Configuration and implementation checks
The table below uses the sample above and lists only settings documented in the official material or implementation checks relevant to this feature and primary task. The hypothetical exercise starting values are chosen for this paper exercise to explain the comparison method; they are not official default values.
| Configuration item | Hypothetical exercise starting value and reason | Checking method |
|---|---|---|
| AI model | A model supporting tool calling and a sufficient maximum context is selected, because the sample contains two long documents | In the AI configuration dialog, check the model's displayed maximum context and tool-calling capability |
| Memory turns | Starting value 2 turns, because the sample is a single-receipt question and does not need earlier history | Compare with the number of history turns actually carried in during a continued session |
| Temperature | A low starting value, because customer-service and rule-defined scenarios need stable answers | Compare answer consistency when the same question is asked repeatedly |
| Reply cap | Starting value set to the length the sample explanation needs, to avoid truncation | Compare with whether the complete answer is truncated early |
| Reply format | Kept at default, because downstream has no fixed-structure reading requirement | If json_schema is used instead, check whether the returned structure matches the filled format requirement |
| Thinking configuration | Kept at default, because the sample question is not complex | Compare answers after switching thinking strength |
Three common errors
- Treating retrieval results as authorisation. In the sample, the receipt voucher and contract being retrieved only means the materials were recalled. It does not mean the trade background has been verified or that settlement has been approved. Retrieval does not perform an authorisation function.
- Treating a similarity score as an entity or permission filter. The score reflects only the degree of recall relevance; it does not exclude unrelated enterprises or unauthorised records.
- Writing a matter still to be verified as already completed. The draft sentence "The trade background of this receipt has been confirmed as authentic, and settlement can be processed" states a conclusion the materials do not support, and exceeds the material scope.
How to verify the setup
- Input the sample text above together with the draft sentence, and compare whether the draft sentence states "confirmed" and "can be processed" as established facts. If the answer instead explains that the materials show only that the payer and payee and the amounts match, and that trade-background authenticity remains to be verified, the condition is met.
- Input a continued-session request with the same chatId, and compare whether the number of history turns actually carried in matches the memory-turns starting value. If the number carried in exceeds the setting, the condition is not met.
- Input the sample text and follow up with "Can this receipt be settled?", and compare whether the answer preserves the evidence gap. If the answer directly gives a conclusion that settlement can proceed, the condition is not met.