What the documents look like
Export tax rebate filing work happens after goods leave the country, when tax vouchers are matched. The material comes from the electronic tax bureau, the export rebate filing system, customs electronic records and the financial ledger. The working documents are the export tax exemption and rebate filing form, the export customs declaration, purchase invoices and foreign exchange receipt vouchers. The data updates by month and by batch.
For order fulfilment and after-sales questions, a user may follow up across several turns about the filing progress of one batch, the match between a customs declaration and a purchase invoice, or how a query feedback item is handled. Each document has a subject, a period and status fields. Those fields decide what a model may quote and what it may not infer. The prompt has to state which documents are provided, which fields are readable, and how to answer when evidence is missing. A model cannot fill in business facts from context.
Constraints at the multi-turn prompts step
Multi-turn dialogue depends on which history is actually carried in. When chatId is passed, the system takes the conversation from the database and uses the last item in the messages array as the user question; the remaining messages are ignored. When chatId is not passed, context is built entirely from the supplied messages. The prompt must state the document scope, the answering task and the handling of missing evidence.
Reply format follows the downstream consumer. Selecting json_schema requires filling in the corresponding format requirement and checking the returned structure. For model and length, check the context the model supports and the reply settings, then test long documents combined with follow-up questions. Question optimisation is checked only when continuous knowledge base conversation is needed, to confirm reference resolution and question expansion.
Hypothetical drill sample: a finance staff member at a foreign trade enterprise asks about the March 2025 batch. The source text reads: "Export customs declaration A-2025-0312, export date 12 March 2025, corresponding purchase invoice B-0312, amount consistent with the customs declaration. The foreign exchange receipt voucher for this batch was registered on 8 April 2025, status registered." The draft sentence to be judged is: "The tax rebate filing for this batch is complete and the rebate has been received." The source text only states that the receipt voucher is registered. It does not state that the filing is complete or that the rebate has arrived. The draft sentence is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| AI model | A model shown in the dialog as supporting the needed context length | Compare the model's stated maximum context with the sample document length |
| Memory turns | Set to 3 turns for this trial | Customer service apps usually keep few history turns; compare whether earlier information is usable at 1 turn and at 5 turns |
| Temperature | A low value for this trial | Rule-defined scenarios need stability; compare answers at a low value and at the default |
| Reply format | Keep default | Ordinary question answering needs no fixed structure; change only if a downstream step reads fixed fields |
| Multimodal recognition | Decide by model capability | The sample is text material; confirm model support before enabling |
| Hide AI output | Keep off | The answer must be shown to the user; adjust temporarily during debugging |
These values are hypothetical trial starting points, not official defaults. The configuration dialog shows available items according to model capability; when a model does not support an item, the corresponding option is not offered.
Three common errors
- Treating history as the whole context. When chatId is passed, the remaining messages are ignored. If the prompt assumes every supplied message takes effect, the answer departs from the actual input. The check compares the messages actually supplied with the last one the system uses.
- Treating retrieval results as an authorisation basis. Knowledge base retrieval returns relevant fragments only; it does not judge business permission or entity ownership. Retrieving a receipt voucher fragment in the sample does not mean the filing is complete. The check compares whether the answer goes beyond the status stated in the source text.
- Treating pending review as approved. The sample source text only states that the receipt voucher is registered; it does not state that a filing was submitted or approved. The check compares each assertion in the draft sentence against the corresponding field in the source text, and marks missing evidence as unsupported.
How to verify the setup
- Input the sample source text and the draft sentence "The tax rebate filing for this batch is complete and the rebate has been received." Compare the assertions "filing is complete" and "rebate has been received" with the source text's "receipt voucher is registered." When the source text does not state the filing status or the arrival of funds, the answer should point out that the sentence is unsupported and keep the evidence gap.
- Input consecutive follow-up questions under the same chatId and compare the number of history turns actually carried in with the configured memory turns. If the configuration is 3 turns, the answer should not cite earlier content beyond that range; content beyond it means the check did not pass.
- Input the sample material and select json_schema format, then compare the returned structure against the format requirement that was filled in. If the returned structure does not match, check whether the format requirement is complete; do not assume the system corrects it automatically.