What the documents look like
A brand concession inside a shopping mall accepts the mall's unified cashiering and commission settlement. The same transaction appears in different forms across the brand ERP, the mall POS, mall member campaign records and the finance system. Operating reports and exception reviews have to place four document types on one chain of judgement: the joint-operation commission contract sets the commission basis, the mall campaign letter sets cost ownership, the sales receipt sets the sales detail, and the monthly settlement statement sets the settled result. Tool calls and plugins matter here because the model has to select which material to read and which fields to extract from context, then hand the result to a downstream node or to manual checking.
The contract and the campaign letter carry different effective windows and different cost rules, so a field taken from one document is not automatically valid under another. The settlement statement arrives monthly, while receipts accumulate daily; a review that mixes periods or subjects produces a comparison that looks complete while resting on mismatched inputs.
Constraints at the tool calls step
A tool introduction states what the tool does, and it also needs to state the input conditions and the situations in which the tool should not be called. Tool parameters divide into fixed parameters for system modules and configurable parameters for plugins. A parameter description is not a guarantee of execution. Where a parameter is marked required and no suitable value exists in context, the model may decline to call the tool; prompt wording can guide the user to supply the missing value. In advanced orchestration, available tools are connected to the tool call module at the diamond connector, and the guidance is to tune a small number of tools before adding more. The tool call termination node ends the current call: no further tool is called and no AI summary of the tool result is produced. Custom tool variables extend the variable inputs of nodes that are not treated as tool parameters. Retrieval and similarity scores carry no permission or entity-isolation duty.
Hypothetical paper exercise sample: the contract text reads "a brand and a mall signed a joint-operation commission contract for the period 1 January 2024 to 31 December 2024, with a commission rate of 20% of the month's tax-inclusive sales." The campaign letter records a mall-wide discount campaign from 1 May 2024 to 5 May 2024, with campaign costs borne by the brand and cost confirmation subject to the campaign settlement statement issued by the mall. The May 2024 monthly settlement statement records tax-inclusive sales of 100,000 yuan, commission cost of 20,000 yuan and campaign cost of 3,000 yuan. The draft sentence under review is "the commission cost for this concession in May 2024 should be calculated on post-campaign sales." The subjects required for judgement are the brand and the mall; the object is the May 2024 commission cost; the state is that both the contract and the campaign letter are within their effective periods. The draft sentence does not match the commission basis wording in the contract text, so it remains unverified, and post-campaign sales cannot be treated as the commission basis on this evidence.
Configuration and implementation checks
The table below lists settings already described in the official material, or implementation check items, tied to this sample and this page's primary task. The hypothetical trial starting point is used only for this paper check and is not an official default.
| Configuration item | Hypothetical trial starting point | How to check |
|---|---|---|
| Tool introduction | State that it reads the joint-operation commission contract, the mall campaign letter and the monthly settlement statement | Compare against the subjects, objects and periods in the source text; check whether the introduction covers them |
| Tool parameters | Plugin parameters set to material type and period range | Check the parameter description and whether it is required; record a non-call when a required parameter is absent |
| Tool connection | Connect only the contract-reading tool and the settlement-statement-reading tool | Confirm the diamond connector in advanced orchestration and observe call order |
| Tool call termination | Placed after the settlement statement reading | Confirm that the call ends without an AI summary of the tool result |
| Custom tool variable | Add a period variable for the campaign letter reading node | Check whether the variable description matches the actual request parameter |
| Prompt guidance | Require the model to ask the user when the period is missing | Observe whether a question is raised when no period is supplied |
Three common errors
- Writing the tool introduction as a general catalogue without the situations in which the tool should not be called. Input: the three material types in the sample. Compare whether the introduction covers the contract period and the campaign period. Observable result: where the introduction omits period conditions, the model may call the tool in an unsuitable period.
- Treating the parameter description as a guarantee of execution. Input: the sample with the campaign settlement statement absent. Compare whether the parameter is required against whether context supplies it. Observable result: where a required parameter is missing and the model still returns a result, record the outcome as unverified.
- Treating a retrieval result as authorization or as a settlement basis. Input: the settlement statement reading result. Compare whether it shares subject, period and field with the contract commission basis. Observable result: where the field basis differs, the value cannot be written directly into the report.
How to verify the setup
- Input the three source texts in the sample and the draft sentence. Compare whether "post-campaign sales" in the draft and "tax-inclusive sales" in the contract text are the same field. Where the fields differ, record the draft as unsupported rather than as a negation.
- Input a request with the campaign settlement statement missing. Compare whether the tool parameter is required against whether context supplies it, and observe whether the model asks a question. Where no question is asked and a result is returned, record the outcome as unverified.
- Input the settlement statement reading result. Compare the commission cost of 20,000 yuan against the relationship between the 20% contract rate and tax-inclusive sales of 100,000 yuan, and observe whether the two agree. Where they agree, this confirms only that the field is checkable within that period; it does not extend to other periods or campaigns.