What the documents look like
This page covers order fulfilment and after-sales question answering during franchisee training and opening coaching. The working materials are training sign-in sheets, opening checklists, SOP manuals and assessment scores. They come from training systems, supervisor apps, store photographs and the head office SOP library, and they are produced per store per month. Opening coaching transfers operating capability; these records show whether an action was completed, whether materials arrived and whether an assessment was passed. That work is separate from franchise filing and disclosure.
Multi-turn dialogue treats these materials as checkable evidence and follows an order or after-sales question across several turns. Because each store produces several sign-in sheets and checklists across different months, a conversation first establishes which store, which month and which document is under discussion before answering.
Constraints at the multi-turn prompts step
The prompt states the document scope, the answering task and how to behave when evidence is absent, so the model does not supply business facts on its own. History handling requires checking what context is actually carried in and distinguishing a new conversation from a continued one. Where a knowledge base supports continuous dialogue, check coreference resolution and question expansion rather than assuming every short question qualifies. Model and length settings require checking the context and reply settings the chosen model supports. Reply format follows the downstream consumer; when json_schema is selected, the format requirement is filled in and the returned structure is verified.
Hypothetical paper exercise: a franchise store submits an opening checklist. The source text reads "Store number A-017, checklist month March 2025, cashier counter materials field 'in place', inventory count field 'pending entry', supervisor signature field blank." A training sign-in sheet for the same period reads "A-017, March 2025, 6 attendees, 6 signatures." The draft sentence to be judged is "Store A-017 completed its March 2025 opening checklist in full, with inventory count and supervisor confirmation both closed out." The draft conflicts with the source at two points, the inventory count field and the supervisor signature field, and must be checked against the source and rewritten. The materials do not state whether the store received opening approval; that fact stays unknown.
Configuration and implementation checks
| Configuration item | Use on this page | How to check |
|---|---|---|
| AI model | Select a dialogue model able to handle long Chinese text | Compare the maximum context and multimodal capability shown in the dialog against the length of the materials in use |
| Memory turns | Control how many earlier turns are carried in | For this paper exercise, start with a small number of turns because the task is mostly single-store, single-month follow-up; compare against carrying in many turns for consumption and response speed |
| Reply limit | Cap the length of a single answer | For this paper exercise, start at a medium length; compare against truncation and observe whether all checklist fields are covered |
| Temperature | Control answer stability | This task has clear rules, so use a low value and compare wording stability against a higher value |
| Reply format | Determine answer structure | Keep the default for ordinary question answering; select json_schema and fill in the format requirement when a downstream node reads a fixed structure |
| Multimodal recognition | Read store photographs | Enable only when the model supports image input; compare against the capability shown in the dialog |
Three common errors
- Treating a retrieval result as an authorization conclusion. The input is a checklist containing "pending entry"; the comparison object is "both closed out" in the draft sentence; the observable result is that the draft writes an unfinished field as finished. Retrieval supplies material and does not constitute approval or closure.
- Conflating a new conversation with a continued one. The input is a follow-up under the same chatId; the comparison object is the number of history turns actually carried in; the observable result is an answer citing old material from another store or month. Check whether chatId and the carried context correspond.
- Treating similarity as an entity filter. The input is a record from another store with a high similarity score; the comparison object is the store number appearing in the answer; the observable result is an answer mixing in records that are not A-017. Similarity does not perform store isolation; verify by store number.
How to verify the setup
- Input the checklist source text and the draft sentence from the hypothetical exercise above. Compare "inventory count and supervisor confirmation both closed out" in the draft against "pending entry" and "signature field blank" in the source. The observable result is that the draft is marked inconsistent with the source, and the answer is rewritten to state the source position and note that supervisor confirmation is unsupported.
- Input consecutive follow-ups under the same chatId. Compare the history turns carried into the answer against the memory turns setting. The observable result is that the carried turns match the setting and the answer cites no material from another store or month.
- Input a query containing only another store number. Compare the store number appearing in the answer against A-017. The observable result is that the answer does not use another store's records as evidence for A-017's fulfilment; if an empty result is returned, check the downstream handling design against the explicit paper input [] and do not claim the query necessarily returns empty.
This page addresses only franchisee training and opening coaching, order fulfilment and after-sales question answering, and multi-turn prompts. At operational handover, which materials, actions and requests sit at which authorization state cannot be treated as authorized execution on the strength of retrieval or a form.