What the documents look like
Replenishment and fulfilment records in this category come from several systems that do not share one owner or one time window: a warehouse management system, an ERP sales forecast, a freight forwarder system, and carrier bills. The working documents are the replenishment plan, the inbound appointment, the inventory age table, and the last-mile delivery note. Updates arrive weekly or daily, so a single question in an after-sales thread can point at a quantity in a replenishment plan, a time slot in an inbound appointment, or a charge line in an inventory age table. Third-party overseas warehouses can serve several sales channels, and the seller allocates stock across them, which means the same SKU can appear in more than one plan with different quantities and dates.
For multi-turn prompts, the practical consequence is that each turn must keep the earlier reference intact and state which document and which field the answer rests on. A reply that moves from a plan quantity to an appointment status without saying so is not traceable. Retrieval results supply reference content only; permissions and business authorization are checked separately and are not established by a similarity score.
Constraints at the multi-turn prompts step
The prompt for this task states the document scope, the answering task, and the handling of missing evidence, so the model does not fill in business facts. History handling is checked against what was actually carried into context, and a new session is distinguished from a continued session. Where a knowledge base supports continuous dialogue, reference resolution and question expansion are checked rather than assumed to apply to every short question. Reply format follows the downstream consumer; selecting json_schema requires filling in the corresponding format requirement and checking the returned structure. Model context length and reply length settings are checked against long documents combined with follow-up questions.
Hypothetical paper exercise: a cross-border seller holds a replenishment plan reading "Plan No. BH-2024-07, subject is a seller in East China, object is US West overseas warehouse SKU-A, planned replenishment 500 units, planned inbound date 10 August 2024, status pending review." A second document, an inbound appointment, reads "Appointment No. YY-2024-08, same subject, same overseas warehouse SKU-A, appointment inbound 500 units, appointment slot 12 August 2024, status submitted." The draft sentence to be judged is "This batch of replenishment has completed inbound, and the last-mile delivery note has been generated." Comparison covers subject, object, quantity, date, and status fields. The plan status is pending review, the appointment status is submitted, and neither document contains a last-mile delivery note record, so the phrases "has completed inbound" and "has been generated" lack supporting evidence.
Configuration and implementation checks
| Configuration item | Value or check for this page | Basis and boundary |
|---|---|---|
| AI model | Select a model whose context length fits the documents | Official description notes differences in context, tool calling, and multimodal ability |
| Memory turns | Hypothetical starting point for this exercise is a small number of history turns | Official description states customer service and knowledge base Q&A applications usually keep a small number of history turns; the specific number is compared against actual dialogue length |
| Reply limit | Set according to required answer length | Official description notes a low setting may truncate and a high setting may increase consumption |
| Temperature | Choose a lower value | Official description recommends lower values for customer service, knowledge base Q&A, and rule-defined scenarios |
| Reply format | Keep default for ordinary Q&A; for fixed structure select json_schema and fill in the format requirement | Official description notes the format is changed only when a downstream process reads fixed-format content |
| Multimodal recognition | Check against model ability and input type | Official description notes that after enabling, requests are filtered again by model ability and unsupported types are not sent |
Three common errors
- Treating a pending-review status as completed. In the hypothetical exercise the replenishment plan status is pending review; if the draft sentence reads "has completed inbound," the input and comparison object is the status field, and the observable result is a mismatch between the status value and the wording.
- Treating a submitted appointment as executed inbound. In the hypothetical exercise the inbound appointment status is submitted; if the draft sentence reads "the last-mile delivery note has been generated," the input and comparison object is the appointment status and the delivery note record, and the observable result is that no delivery note field exists in the material.
- Treating a retrieval result as an authorization basis. In the hypothetical exercise retrieval returns a replenishment plan fragment; if that is taken as grounds to execute an inbound operation, the input and comparison object is the retrieval output and the authorization material, and the observable result is that the retrieval output contains no approval or execution field.
How to verify the setup
- Input the two document texts and the draft sentence from the hypothetical exercise, compare subject, object, quantity, date, and status fields, and treat the draft as satisfied when it matches the material status; record it as unsupported when it does not.
- Input the hypothetical replenishment plan status of pending review, compare whether the draft sentence contains a "completed" type expression, and mark it as beyond the evidence scope when such wording appears.
- Input the hypothetical inbound appointment status of submitted, compare whether the draft sentence contains a "generated" type expression, and mark it as missing the corresponding record when such wording appears.