What the documents look like
Near-expiry markdown handling works on shelf stock batches rather than on weighed, sorted order lines. The working records are the near-expiry list, markdown approval, damage write-off form, and promotional price tag. Supporting data comes from POS stock age, store inspection, ERP inventory, and quality inspection records, refreshed daily or per shift. Store service and internal training must let the store manager, shrinkage administrator, and finance auditor obtain checkable answers across a multi-turn conversation. Prompt text therefore has to state the material scope, the answering task, and the treatment of missing evidence.
Two features of this category shape the multi-turn step. First, the object of every answer is a batch with a shelf-life date and an approval status, so a reply that mixes batches or periods is unusable. Second, approval, write-off, and tag placement are separate business stages, and a printed tag is not a placed tag. A prompt that lets the model fill these gaps produces answers that look complete while resting on nothing.
Constraints at the multi-turn prompts step
Multi-turn dialogue carries earlier turns into the request. A new session and a continued session accept different inputs, so the actual context that reaches the model must be checked rather than assumed. Memory turns control how many earlier rounds the model may reference; a small value can leave earlier information unused. Question optimization applies to continuous knowledge-base dialogue and covers pronoun resolution and question expansion; short questions do not automatically qualify. The model's maximum context sets the length of material that can be referenced, and the reply cap controls a single answer's length. Reply format follows what the downstream step reads; choosing json_schema requires filling in the corresponding format requirement and checking the returned structure. Retrieval returns candidate content and does not constitute authorization. Forms collect values, while business verification, approval, database writes, and external synchronization each need their own downstream implementation.
Hypothetical drill sample: the handover material for checking consists of three sentences. Sentence one: "Shift inspection record, 10 March 2026: cold cabinet A, batch 12, low-temperature yogurt, shelf life to 12 March 2026, status pending markdown approval." Sentence two: "The markdown approval form for this batch has not been signed; the approver field is empty." Sentence three: "The promotional price tag has been printed and has not been placed on the shelf." The draft sentence to be judged is: "This batch has completed markdown approval and may be sold at the new price tag." The subject is this batch, the object is the markdown approval status, the period is 10 March 2026, and the field is the approval signature status. The source states only that the approval form is unsigned, while the draft sentence states that approval is complete. Same subject, same period, same field, mutually exclusive values: the draft is unsupported.
Configuration and implementation checks
The table lists only settings documented for this function and this task, or implementation checks. The numbers are starting points for this hypothetical drill, with the reason for choosing them and the value they are compared against.
| Configuration item | Starting point for this hypothetical drill | Reason and comparison method |
|---|---|---|
| AI model | Choose a model whose displayed maximum context fits the material | Compare the model's displayed maximum context with the combined length of the sample material and the follow-up question |
| Memory turns | Keep a small number of history turns | Compare with the number of earlier turns the sample needs, to confirm the turn holding the approval status is carried in |
| Reply cap | Turn on and set to hold a complete explanation | Compare with the length of the sample draft sentence to confirm it is not truncated early |
| Temperature | Low | Compare with the need for stable answers in a rule-explicit setting |
| Reply format | Keep default | No change to json_schema is needed when the downstream step does not read a fixed structure |
| Hide AI output | Keep off | The sample needs the reply shown directly for checking |
Three common errors
- Treating a retrieval candidate as an authorization basis. Input: the sample batch and the approval status query. Comparison: the retrieval return against the approval signature field. Observable result: the reply infers completed approval from a retrieval hit. Retrieval is candidate content; the approval status must be checked against the approval form field.
- Treating a form value as business verification. Input: the value entered in the markdown approval form. Comparison: the entered value against the approver signature field. Observable result: the reply treats entry as signature. A form collects values; approval signature is a separate fact.
- Treating pending review as missing permission. Input: the unsigned status of the sample approval form. Comparison: the unsigned status against the operation permission field. Observable result: the reply derives a lack of permission from the lack of signature. Pending review does not show a lack of operation permission.
How to verify the setup
- Input the three sample sentences and the draft sentence. Compare the draft phrase "has completed markdown approval" with the source phrase "the approval form has not been signed." The condition is met when the reply states that the same subject, same period, and same field hold mutually exclusive values and judges the draft unsupported.
- Input the sample batch query request. Compare whether the reply treats a retrieval hit as completed approval. The condition is met when the reply keeps the approval status as pending verification and states that the approval form signature field must be checked.
- Input the sample promotional price tag state, "printed and has not been placed on the shelf." Compare whether the reply states that sale at the new price tag may proceed. The condition is met when the reply states that placement and sale are unexecuted actions and preserves the evidence gap.