What the documents look like
Opening coaching is a transfer of operating capability, not a legal filing disclosure. The materials that multi-turn dialogue must carry are training sign-in sheets, opening checklists, SOP manuals and assessment results, refreshed per store and per month. A supervisor may first ask whether a given store's sign-in records are complete, then ask which session is missing, then ask for a retraining suggestion that references the SOP manual. The prompt must state the document scope, the answering task and the handling of missing evidence so that the model does not fill in business facts that were never supplied. History determines whether a follow-up question can reuse earlier context, and a new session differs from a continued session in what context it carries in.
Constraints at the multi-turn prompts step
The memory turn count controls how much prior dialogue the model can reference. A small value drops the store object; a large value pulls in unrelated content. The reply cap affects how complete an SOP action description can be, and truncation can drop a key step. A lower temperature suits rule-based checking, while a higher value suits drafting retraining plans. Reply format follows downstream consumption: change it only when a later step reads a fixed structure, and when json_schema is selected, the corresponding format requirement must be filled in and the returned structure checked. Retrieval results are candidate materials only. They do not constitute authorization and do not replace human verification.
Hypothetical paper exercise sample: a franchise store's June 2025 training sign-in sheet records "June 3 cashier training, 4 expected, 3 present, 1 absent with no reason noted"; the opening checklist records "June 5 inspection, cashier counter materials complete, inventory count not finished"; SOP manual Chapter 3 requires "all staff present at cashier training before taking a post." The draft sentence to be assessed is "This store has completed all opening training and can be scheduled to take a post."
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison object |
|---|---|---|
| Memory turns | 4 turns | Covers the four follow-ups on sign-in, absence, materials and inventory; compared with 2 turns to see whether the store object is lost |
| Reply cap | Enabled at a medium length | Holds an SOP step description; compared with a lower value to see whether steps are truncated |
| Temperature | Lower | Checking answers need stability; compared with a higher value to see whether wording drifts |
| Top_p | Left off or at default | Stability already obtained through temperature; no added randomness |
| Reply format | Default | No downstream step reads a fixed structure; change only when fields must be read |
| Multimodal recognition | Image enabled per model capability | Needed for store photo checks; the option is not shown when the model does not support it |
These values are a hypothetical starting point for this exercise and are not official defaults. The model and length must be checked against the context and reply settings the model supports, and long-document plus follow-up combinations must be tested.
Three common errors
- Treating a retrieved sign-in sheet as a post authorization. Retrieval returns candidate materials only; authorization status must be checked separately. The error appears when the draft sentence skips the absence fact and states a post-taking conclusion.
- Treating an opening checklist still under review as completed. The checklist records that inventory count is not finished, which is an incomplete state. The error appears when the draft sentence says "completed all opening training."
- Mixing records from different stores or different periods into one follow-up. The sign-in sheet is dated June 3 and the checklist June 5, both for the same store in the same period; substituting another store's record breaks the subject match. The error appears as a retraining suggestion attached to the wrong store.
How to verify the setup
- Input the sign-in sheet text above and the draft sentence. Compare whether the draft sentence agrees with "3 present, 1 absent." Observe whether the model points out that the absence has no reason noted and that training completion cannot be determined from this record.
- Input the opening checklist text above and the draft sentence. Compare "inventory count not finished" with "completed all opening training" as mutually exclusive values of the same field. Observe whether the model marks the claim as unsupported and keeps the evidence gap open.
- Input a continued-session follow-up, "which session is missing," and compare whether the returned content stays with the same store and the same period. Observe whether the store object is retained under the memory turn setting; if another store's record is returned, record it as context loss.