What the documents look like
Export tax rebate filing work happens after goods leave the country, and it matches tax vouchers against what was declared. The main data sources are the electronic tax bureau, the export rebate filing system, customs electronic messages and the finance ledger. The documents handled day to day are the export tax exemption and rebate filing form, the export customs declaration, purchase invoices and foreign exchange receipt vouchers. Mistakes in this category produce delayed rebates, ineligible rebates or tax risk, so the input material for copy drafting has to stay inside filing records that can be checked line by line.
The task on this page is product and marketing copy drafting. The multi-turn prompts step has to settle three things inside the prompt: the scope of the material, the drafting task, and how to behave when evidence is missing. The model drafts copy from the material that exists, and any promotional claim the material does not cover stays as a gap. History turns, question optimisation, model and length settings, and response format all affect whether the drafted copy can be checked against the source.
Constraints at the multi-turn prompts step
Multi-turn dialogue carries earlier turns into the current request. If an unchecked promotional phrase appeared in an earlier turn, a later turn can reuse it. Checking what context actually entered the request, and separating a new session from a continued session, is a required action. Question optimisation is used in knowledge base continuous dialogue for reference resolution and question expansion, and whether it applies to a short question depends on the actual input. The context length a model supports and the reply cap together decide whether a long document combined with follow-up questions can be handled completely. Response format follows how the output is consumed downstream; when json_schema is selected, the format requirement is filled in and the returned structure is checked.
Hypothetical paper exercise: a finance staff member at a foreign trade enterprise assembles filing material for one batch of exported goods. The material text is three sentences: "This batch of goods was exported in March 2025, and the commodity name on the customs declaration matches the purchase invoice. The purchase invoice states a quantity of 1000 pieces, and the foreign exchange receipt voucher shows the payment for this batch has been received. The export tax exemption and rebate filing form has been completed with this batch's data, and its status is pending review." The draft sentence to be judged is: "This product has passed export rebate review and can be purchased with confidence." The comparison object is "has passed export rebate review" in the draft against "status is pending review" in the material. Under the same entity, the same batch of goods and the same filing field, "pending review" and "passed review" are mutually exclusive values for one field, so the draft sentence contradicts the material. The material states no review conclusion, so that promotional claim in the draft is unsupported.
Configuration and implementation checks
The configuration below centres on that exercise and lists only settings the official documentation describes, or implementation checks, that relate to this feature and this task. The values are starting points for this hypothetical trial, with the reason for choosing them and the value they are compared against.
| Configuration item | Hypothetical starting point | Reason and comparison method |
|---|---|---|
| Memory turns | Keep a small number of history turns | Drafting needs only recent material; compare against carrying all history turns and observe whether the draft reuses unchecked phrasing |
| Reply cap | Set according to the drafting length needed | Copy needs complete sentences; compare against a setting that is too low and observe whether the reply is cut off early |
| Temperature | Low | Drafting relies on material, and stable output is easier to check sentence by sentence; compare against a higher setting and observe whether wording drifts from the material |
| Response format | Chosen by downstream consumption | When a later flow reads a fixed structure, select json_schema and fill in the format requirement; compare the returned structure against the default format |
| Prompt | State material scope, drafting task and missing-evidence handling | Compare against a prompt without missing-evidence handling and observe whether the draft adds a review conclusion the material does not state |
| Hide AI output | Toggled as debugging needs | Temporarily turn it off while debugging to observe intermediate content; compare against the on state and confirm the final reply can still pass to later nodes |
Three common errors
- Writing "passed review" from material that says "pending review". The input is the filing status field in the material and the draft sentence; the comparison is the value of the same field; the observable result is that the two values are mutually exclusive. When evidence is missing, the gap stays and is recorded as unsupported.
- Using unchecked promotional phrasing from a history turn as this turn's basis. The input is the history context that actually entered and the current draft sentence; the comparison is whether the draft sentence has matching source text in this turn's material; the observable result is that phrasing with no matching source text is marked as needing verification.
- Treating a similarity score as authorisation or as an entity filter. The input is the similarity returned by retrieval and the record identifiers; the comparison is the entity that owns the record against the target entity; the observable result is that a high similarity does not change the record's authorisation ownership, and retrieval does not perform authorisation judgement.
How to verify the setup
- Use the three material sentences from the hypothetical exercise and the draft sentence "This product has passed export rebate review and can be purchased with confidence." as input; compare the draft sentence against the filing status field in the material; the expected result is that "passed review" and "pending review" are identified as mutually exclusive and that claim in the draft is unsupported.
- Use the empty result input [] to check the downstream design. The input is an empty quote list; compare the downstream node's output when there are no quotes; the expected result is that downstream does not add a review conclusion the material does not state, and the gap remains unsupported.
- Use the material text for the same batch of goods and one handover request to check whether the prompt states how missing evidence is handled. The input is the material text and the request; compare the prompt content against the judgement basis the request needs; the expected result is that the prompt clearly requires promotional claims not covered by the material to stay as gaps, and that manual checking actions and business verification actions are recorded separately.