What the documents look like
Price and display auditing happens during routine operations, and its focus is sustained execution of standards rather than launch support. The working documents are store visit photographs, price tag lists, display standards and rectification notices. Source records update weekly or monthly, and that cadence shapes how much history a multi-turn drafting session can carry: a price tag conclusion from the previous visit may still be relevant to the current draft, or it may have lapsed once the promotion period changed. Supervision systems, POS price tables, store cameras and head office promotion schedules supply the facts a draft may rest on. Drafting stays inside what those sources actually state; promotion intensity, price validity windows and rectification completion status are not filled in.
Constraints at the multi-turn prompts step
Multi-turn drafting requires an explicit statement of the material scope, the answering task and the treatment of missing evidence. History is checked against the context actually carried into the request, and a new session is distinguished from a continued one. Question optimization is checked only where a knowledge base conversation needs pronoun resolution and question expansion. Model and length settings are checked against the context and reply configuration the selected model supports, with long documents and follow-up combinations tested together. Reply format follows the downstream consumer; choosing json_schema requires the corresponding format requirement to be filled in and the returned structure verified.
Hypothetical paper exercise: a regional supervisor submits the following store visit material in a paper walkthrough — "March 12 store visit record: the end-cap price tag at Store A shows 9.9 yuan, and the POS price table for the same date lists the item at 9.9 yuan; the head office promotion schedule marks this promotion as running March 10 to March 20 at 9.9 yuan." The draft sentence to be judged is: "Store A executed the price correctly, and the price was consistent throughout the promotion period." The judgment requires the same subject, the same period and the same field. The material covers a single day, March 12; the promotion schedule covers March 10 to March 20. The draft sentence extends a single-day check into a whole-promotion-period conclusion, which the material does not support. If the material states only that the price tag and the POS agree and does not include the promotion schedule, the promotion-period conclusion remains unverified.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| AI model | A model that supports multimodal recognition is selected | Store visit photographs need image recognition; compare recognition capability against a model without multimodal support |
| Memory rounds | Assumed starting point of 4 rounds | Covers the previous store visit and the current follow-up; compare with 2 rounds to see whether the earlier price tag conclusion is lost |
| Reply limit | Assumed starting point of 800 | Copy drafting needs complete paragraphs; compare with 300 to see whether output is truncated |
| Temperature | Assumed starting point on the low side | Price and display wording needs stability; compare with a higher value for wording drift |
| Reply format | Default text; switched to json_schema when a downstream node reads fixed fields | Where a downstream node needs fixed fields, fill in the format requirement and verify the returned structure |
| Multimodal recognition | Image recognition enabled | Store visit photographs are image input; compare with the disabled state to see whether image content is read |
These values are hypothetical starting points for this exercise, not official defaults. The AI configuration dialog shows only the options the current model supports; where a model does not support multimodal recognition, no such option appears, and where it does not support thinking configuration, no thinking option appears. Memory rounds control how many earlier chat rounds the answer may reference, and a larger value carries more content, which can raise consumption and affect response speed. The reply limit caps a single answer, and a low setting can cut a reply short. Temperature controls stability, and low values suit customer service, knowledge base question answering and rule-bound scenarios. Topp overlaps with temperature, and adjusting both at once is not recommended. Stop sequences end output at a fixed marker and are not needed for ordinary chat. Reply format stays at default for ordinary chat, customer service and knowledge base question answering, and jsonschema suits cases where the model must return a fixed structure. Multimodal recognition depends on the model itself, and even when a type is enabled the request is filtered again before sending. Media links in a user question are parsed only when link extraction of multimodal files is enabled, only when the question is under 500 characters, and at most 4 media links are processed at a time. Ordinary document files are not sent as multimodal input; their content still needs file parsing into text. Hiding AI output keeps generated content from being shown directly while still passing it to later nodes.
Three common errors
- Writing a single-day check as a whole-promotion-period conclusion. The input is the March 12 store visit material; the comparison object is the phrase "throughout the promotion period" in the draft sentence; the observable result is that the draft sentence covers a period wider than the material.
- Writing a pending-review state as completed rectification. The input is a rectification notice in pending-review state; the comparison object is the draft sentence "rectification is complete"; the observable result is that the draft sentence treats pending review as completion.
- Treating retrieval relevance as a permission filter. The input is similar items returned by knowledge base retrieval; the comparison object is the draft sentence "it is confirmed that this store is authorized to view this"; the observable result is that the draft sentence uses similarity as a basis for authorization.
How to verify the setup
- Input the March 12 store visit material and the draft sentence from the hypothetical exercise; compare the period covered by the draft conclusion with the period covered by the material; observe whether the draft sentence extends a single-day conclusion into a promotion-period conclusion. The expected result is that the draft sentence is marked as unsupported.
- Input the paper record showing the rectification notice in pending-review state and the draft sentence "rectification is complete"; compare the state field values; observe whether the draft sentence treats pending review as completion. The expected result is that the draft sentence is marked as unverified.
- Input the paper empty result [] together with the downstream handling design; compare whether the downstream branch handles the empty result; observe whether the downstream treats an empty result as a populated one. The expected result is that the downstream follows the empty-result branch, without assuming retrieval must return empty.
All three verifications are paper expectations and were not actually executed. A relevance score is not a category, entity or permission filter, and retrieval does not confer authorization. Where a document field is not provided, the value remains unknown.