What the documents look like
Near-expiry markdown handling tracks the shelf-life lifecycle of store inventory. It does not weigh orders item by item the way sorting operations do. Store inspections, ERP inventory records, and quality inspection records together determine whether a batch can be discounted, the discount depth, and the remaining sale window. These records carry day-level and shift-level update frequency, so a batch state cited in draft copy can change after the draft is written.
Copy drafting for goods and marketing therefore works from batch materials that have explicit time boundaries. The discount reason, the remaining sale period, and the quality description in a draft sentence must each be traceable to a field in the near-expiry list, the markdown approval, the write-off form, or the promotional price tag. A model generates candidate text from the materials supplied to it. It does not decide batch status. Where the materials are silent on a discount reason, a quality conclusion, or an authorization boundary, the model output cannot supply that content.
Constraints at the model connections step
The model reads the text content of the near-expiry list, markdown approval, write-off form, and promotional price tag, and returns candidate copy. Application-side parameters are set in the orchestration node. A compatible chat interface field with the same name does not override the orchestration setting, and this should not be assumed.
AI Proxy is a required dependency. Both the App service and the Admin service must set a root address and a non-empty administrator token. The root address is the AI Proxy service root and does not carry a /v1 suffix. The administrator token must match the server configuration and cannot be replaced by an application API key. A language model must be enabled for the text drafting input. The prompt input range covers only the four document types named above; the input node references those document fields. Model call logs retain request records for the documented retention period, and the log page shows request records.
Assumed drill sample. On 12 March 2026, a store's morning-shift inspection record states: "Batch A yogurt, shelf life to 15 March, outer packaging intact, entered on today's near-expiry list." The same day's markdown approval states: "Batch A, markdown depth 7 zhe, must be sold before 20:00 today, may not carry over to the next day." A draft sentence under review reads: "This batch of yogurt is discounted because it is near expiry, quality is as usual, safe to buy through 16 March."
The materials support the markdown and the same-day deadline. They do not state a quality inspection conclusion of "quality as usual," and they do not authorize an extension to 16 March.
Configuration and implementation checks
| Configuration item | Assumed trial starting point and reason | Check method |
|---|---|---|
| AI Proxy root address | Fill the service root address without appending /v1; the documented path form does not include /v1 | Compare the written path against the deployment document |
| AI Proxy administrator token | Use a non-empty value consistent with the server side; the token must match and cannot be an application API key | Compare the two configuration strings |
| Model capability | Confirm against the text input the drafting task requires | Confirm item by item that a language model is enabled |
| Application parameters | Set in the application or node configuration | Check the parameter entries inside the orchestration |
| Prompt input range | Include only the list, approval, write-off, and price tag text | Inspect the document fields referenced by the input node |
| Model call log retention | Check against the documented retention duration | Confirm request records are visible on the log page |
Three common errors
- Treating a quality conclusion absent from the materials as supported. The sample materials state only that the outer packaging is intact. The draft sentence says "quality is as usual," which the materials do not support.
- Expanding the same-day deadline in the markdown approval into a longer sale period. The approval states before 20:00 today; the draft sentence says through 16 March, which exceeds the authorized range.
- Treating model-generated copy as already passing business review. Model output is candidate text. Markdown approval and price tag release require downstream implementation.
How to verify the setup
- Input the sample material text and the draft sentence. Compare whether the quality description in the draft sentence has a corresponding field in the materials. The materials contain only "outer packaging intact" and no quality inspection conclusion, so the sentence is judged unsupported.
- Input the markdown approval text and the draft sentence. Compare the remaining sale period field. The approval gives before 20:00 today; the draft sentence gives 16 March. The two values for the same field are inconsistent, so the sentence is judged to exceed the authorized range.
- Input a paper material containing only a batch number and a markdown depth. Observe whether the model output contains quality or period statements that do not appear in the material. If such statements appear, record them as generation beyond the material range and require human interception.
A pending request is not approval. An offered option is not execution. An unsigned termination is not a completed termination. Where a field is not provided, it remains unknown.