What the documents look like
Permit handling happens before customs declaration, and it determines whether a shipment may lawfully cross the border. That timing difference means the materials that appear inside a conversation are usually still under review rather than final. Business records come from commerce authority systems, customs commodity codes, supplier files and customer end-use statements. Storefront service and internal training work most often with automatic import licences, dual-use item licences, contracts and product descriptions. These records update irregularly, on a per-shipment basis, and one shipment can move from an enquiry stage into a pending-supplement stage and then into a pending-review stage across several turns. A multi-turn prompt therefore needs to keep the material version attached to each question, so a later supplementary note does not overwrite an earlier original end-use description.
Two facts shape this page. First, the permit decision precedes declaration, so a conversation about a shipment may sit entirely in an unapproved state. Second, the same shipment record can be revised between turns, which makes version tracking a practical requirement rather than a formatting preference.
Constraints at the multi-turn prompts step
The prompt must state the scope of materials, the answering task and the handling of missing evidence, and it must not write business facts on the model's behalf. History records need a check of the context actually carried in, with new sessions distinguished from continued sessions. Where continuous knowledge-base dialogue is required, check coreference resolution and question expansion, and do not assume every short question fits that pattern. Model context and reply length follow the configuration, and long documents combined with follow-up questions deserve separate observation. Reply format follows the downstream step; when json_schema is selected, fill in the corresponding format requirement and verify the returned structure.
Hypothetical drill sample: a storefront receives a product description whose original text reads, "This product is used for industrial cleaning equipment, the export destination is Southeast Asia, accompanied by automatic import licence application number A-2024-001; the licence is still under review and no certificate has been obtained." The contract clause for the same shipment reads, "The seller shall ship within 30 days after obtaining the licence." A draft sentence awaiting review reads, "This shipment already holds an automatic import licence and shipment may be arranged as agreed in the contract." The check compares the draft sentence against "still under review and no certificate has been obtained" in the product description, and against the sequencing condition "after obtaining the licence" in the contract clause, and concludes that the draft sentence lacks evidential support.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relation to the sample |
|---|---|---|
| Prompt | State material scope, answering task and handling of missing evidence | The prompt requires that an unobtained certificate status is not written as already held |
| Memory turns | Check the number of context turns actually carried in | Preserve the first-turn product description text for later follow-up comparison |
| Reply limit | Set according to reply length needs | Avoid premature truncation when the end-use description is long |
| Temperature | Use a lower value for rule-clear scenarios | Reduce divergent wording about licence status |
| Reply format | Choose according to the downstream step; fill in the format requirement when json_schema is selected | Fix the structure when a downstream step reads a status field |
| Multimodal recognition | Enable the corresponding type as needed when the model supports it | An image scan of a licence requires the model to have image recognition capability |
Three common errors
- Writing "still under review and no certificate has been obtained" as "already holds a licence". The input is the original product description text; the comparison is between the draft sentence and the status field in the original text; the observable result is whether the draft sentence contains an obtained-licence statement the original text does not carry.
- Treating a pending-review status as granted authorisation and then arranging shipment. The input is the contract clause "ship within 30 days after obtaining the licence"; the comparison is between the current status and the clause precondition; the observable result is whether a shippable conclusion is given while no certificate has been obtained.
- Overwriting the first-turn end-use description with a later supplementary note in a continued session. The input is two turns of messages under the same chatId; the comparison is between the first-turn end-use description and the later content; the observable result is whether the first-turn information remains retrievable in the context.
How to verify the setup
- Enter the hypothetical sample product description text together with the draft sentence "This shipment already holds an automatic import licence and shipment may be arranged as agreed in the contract". Compare the draft sentence against "still under review and no certificate has been obtained". The condition is satisfied when the reply points out the missing obtained-licence evidence and keeps the pending-verification status.
- Enter the contract clause text together with a shipment request. Compare the request point in time against the "after obtaining the licence" condition. The condition is satisfied when the reply points out that no certificate has been obtained at present and the shipment condition is not yet met.
- Under the same chatId, first enter the product description and then enter a follow-up question. Check whether the first-turn end-use description remains retrievable in the history record. If the first-turn content is overwritten by the later turn and becomes invisible, record this as a context-carrying mismatch, adjust the memory turns setting and recheck.
When retrieval returns an empty result, an explicit paper-input empty list may be used to check the downstream design, without claiming that a fictional rule search will certainly return empty. A similarity score does not act as a permission or entity filter, and retrieval does not carry authorisation judgement. Libraries spanning different vector models cannot be jointly associated; check the identifier first and exclude non-matching ones.