What the documents look like
Price and display auditing takes place during day-to-day operations, where the focus is sustained execution of standards rather than onboarding or launch support. The working materials come from the supervision system, POS price tables, store cameras and head-office promotion calendars. The documents handled day to day include store-visit photographs, price-tag lists, display standards and rectification notices. This page addresses one narrow step: using store-visit photographs, price-tag lists and similar materials as reference context when drafting price-tag notes, promotion notices or rectification communication copy. The materials are refreshed weekly or monthly. No batch synchronization, full synchronization or real-time trigger mechanism follows from that cadence here.
Two features of this material shape the drafting step. First, price and promotion facts sit in more than one field. A price-tag list records the tag price; a promotion calendar records a campaign price with its own date range. A drafted sentence that names one field does not thereby state the other. Second, the same product can appear across several stores and several dates, so a draft must keep subject, period and field aligned before any claim about execution is treated as supported.
Constraints at the context limits step
The maximum context of a model sets how much content a single request can reference. The reply limit only caps output length. The two are separate settings and cannot be treated as one. The knowledge-base reference limit is controlled by tokens, so a long single passage may be truncated; this needs checking against a real question sample. History turns, question optimization and reranking each change what context reaches the model. Retrieval similarity affects recall ordering only. It does not perform permission filtering or entity isolation, and it does not exclude unrelated stores or unauthorized records. The association between a knowledge base and a vector model is checked against the documented identifiers, not assumed.
Hypothetical paper test case: the source material reads, "The store visit date in this hypothetical case is 10 March 2025. The price-tag list for Store A shows a hypothetical price of 2 yuan for a 500 ml bottled drinking water. The head-office promotion calendar states that this item runs a hypothetical promotional price of 1.8 yuan from 1 March to 15 March 2025." The draft sentence to be judged is: "On the hypothetical date of 10 March 2025, this store executed the promotional price of 1.8 yuan for this item." Judging requires the same subject, the same period and the same field. The material states two fields, the tag price and the campaign price. The draft cites only the campaign field and does not state the actual execution status on that day, so it remains unverified.
Configuration and implementation checks
| Configuration or check item | Use on this page | How to check |
|---|---|---|
| AI model and maximum context | Carries the source material and the draft sentence | Shown according to model capability; a larger value suits longer material |
| Reply limit | Caps the length of the drafted sentence | Limited by a slider once enabled; too low a value truncates |
| Knowledge-base reference limit | Controls referenced content by token | Checked against single-passage length and the sample for truncation |
| Search mode | Semantic, full-text or hybrid recall of material | Numbered or coded content suits full-text or hybrid retrieval |
| Minimum relevance | Filters low-relevance results | Applies only under semantic retrieval or reranking |
| Reply format | Carries draft fields in a fixed structure | When json_schema is selected, the format requirement is filled in |
A hypothetical starting value for this page: the reference limit is set to a token value large enough to hold the three sentences of material above. The reason for choosing it is that the material contains a date, a subject and two price fields, and it needs comparison against a smaller value to observe whether the basis of the draft sentence is truncated. This value is a hypothetical starting point, not an official default.
Three common errors
- Treating the reply limit as the full context window. The reply limit caps output length only; whether input material exceeds the maximum context is a separate check. Check: input the material above, compare the model's referenceable content length against the material length, and observe whether the material enters complete.
- Using a similarity score in place of permission or entity isolation. The score affects recall ordering only and does not exclude unrelated stores or unauthorized records. Check: input mixed material covering Store A and Store B, compare the recall result against the authorized scope, and observe whether out-of-scope content appears.
- Writing a campaign price as the price executed that day. The material states only the campaign period and the tag price; it does not state actual execution on the day. Check: take the draft sentence above as input, compare its fields against the material fields, and observe whether an unstated execution status is present.
How to verify the setup
- Input the material above together with the draft sentence, compare whether the draft cites only fields stated in the material, and observe whether unstated content such as same-day execution or supply period appears. Its appearance means the draft is unsupported.
- Use a clearly labelled paper input of an empty result to check the downstream handling design. Compare how the downstream stage handles an empty reference and observe whether it returns a prompt as designed. Do not claim that a fabricated rule search necessarily returns an empty result.
- Input the same question once with empty history and once with history, compare the two contexts that reach the model, and observe whether the history-turn setting changes the referenced content. Record the result as an actual observation.
All judgments above are limited to the hypothetical sample. Fields that are unstated, or supported only by evidence of a related stage, are recorded as unverified, and the evidence gap is preserved.