What the documents look like
Closure accounting runs in the opposite direction from site selection and store opening. It handles exit and asset recovery rather than launch and expansion. The working materials come from POS inventory, the fixed-asset system, lease contracts and financial ledgers, and the everyday documents are closure approvals, inventory clearance sheets, lease termination agreements and asset count sheets. These records are produced per store and per stage, updated at irregular intervals, and kept separate for each store.
For multi-turn prompts in store service and internal training, staff and trainees ask follow-up questions about those documents, and the model needs to cite the correct material fragment and state the gap when evidence is missing. The stage status of a record directly limits what an answer may claim. An approval that is pending review, an agreement that is unsigned, and a quotation that is within its validity period cannot be treated as completed facts.
Constraints at the multi-turn prompts step
Multi-turn context comes from the history that is actually carried in. A new session and a continued session accept different input: when chatId is passed, the system takes the session from the database, uses the last item of the messages array as the user question, and ignores the remaining messages. The prompt needs to state the material scope, the answering task and the handling of missing evidence, so the model does not write in business facts on its own.
A relevance score in retrieval results is a similarity value. It does not act as a permission filter or an entity filter, and retrieval by itself is not authorization. Forms collect values; business verification, approval, database writes and external synchronization each need their own downstream implementation. An interactive node returns a display configuration containing only type and params, while internal runtime fields are not returned. To continue running, chatId must stay the same, and form input is sent as an object serialized to a string in messages. Thinking budget and the reference length limit are different concepts and cannot be substituted for each other.
Configuration and implementation checks
Hypothetical paper exercise: a chain-brand operations staff member submits an internal-training handover check request. The material text supplied for the check reads: "Closure approval form (hypothetical number BS-001) current status is pending review. Lease termination agreement (hypothetical number ZJ-001) signature status: unsigned by both parties. Asset count sheet (hypothetical number PD-001) record date: hypothetical 1 March 2025, count scope: all fixed assets of this store." The draft sentence to be judged reads: "The store closure approval has passed, the lease has been terminated, and the assets have been handed over." The configuration items below are checked against this sample.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| AI model | A model whose supported context length fits the sample material | Compare against the length of the sample material to confirm the model's maximum context can hold the material plus follow-up questions |
| Memory turns | Hypothetical starting point of a small number of history turns | Compare against the range of prior text the training scenario needs, to confirm the previous turn's material can be cited |
| Reply limit | Hypothetical starting point long enough to hold a complete explanation | Compare against the length of the draft sentence to confirm it is not truncated early |
| Temperature | Hypothetical starting point on the low side | Compare against rule-clear scenarios to confirm stable answers |
| Reply format | Keep the default | Compare against downstream consumption needs; leave unchanged when no fixed structure is read |
| Hide AI output | Keep off | Compare against the training display need; turn off when answers must be shown to trainees |
Three common errors
- Treating pending review as passed. In the sample the closure approval form status is pending review, while the draft sentence writes "has passed." These are mutually exclusive values of the same field being misused. During the check, compare the approval form status field with the wording of the draft sentence.
- Treating unsigned as terminated. In the sample the lease termination agreement signature status is unsigned by both parties, while the draft sentence writes "the lease has been terminated." Compare the signature status with the draft sentence wording; unsigned does not constitute a completed termination.
- Treating a count record as a completed handover. In the sample the asset count sheet has only a record date and a scope, while the draft sentence writes "the assets have been handed over." Compare the count sheet fields with the draft sentence wording; a record date does not equal a completed handover action.
How to verify the setup
- Input the sample material text and the draft sentence, and compare whether the draft sentence matches the material's status fields. If the draft sentence contains "has passed," "has been terminated" or "has been handed over" while the material is respectively pending review, unsigned, or only a record date, then the draft sentence is judged unsupported by the material and must be revised to wording consistent with the status.
- Input a follow-up question, "What stage is this store's approval at now," and compare whether the answer cites the approval form status field. If the answer contains "has passed" or "completed," it is inconsistent with the material's "pending review" and the check fails; if the answer is "pending review" or states that the evidence is insufficient, the condition is met.
- Input an empty-result downstream scenario, using the explicit paper input
[]to check the downstream design. Compare the downstream node's output when it receives an empty reference. If the downstream node fabricates material content directly, the check fails; if the downstream node states that no relevant content was retrieved or routes to manual review, the condition is met. The retrieval check records only observations and failure branches, and does not guarantee a first hit or full recall.
All of the above is a paper check. No test, deployment or approval action was actually executed.