What the documents look like
Promotion enrollment and price management turn on price, inventory and campaign timing. The enrollment form records enrolled products, campaign prices and effective windows. The promotional price approval form records approval status and the authorization scope. The inventory lock table records locked quantities and release conditions. The campaign review table records actual transactions and exception notes. These documents differ in subject, object, status and validity scope, so any cross-document comparison starts by confirming that the fields refer to the same subject and the same period.
Data sources include the Tmall marketing backend, the JD campaign backend, ERP inventory and historical transaction prices. Document update frequency is weekly and per campaign. Tool calls and plugins let the model decide from context whether to invoke a module or plugin, which dynamically arranges report checking and exception review. The tool description, tool parameters and whether a parameter is required together determine invocation behavior. When a required parameter has no suitable value in context, the model may not call the tool.
Constraints at the tool calls step
Tool calls suit dynamic decisions, at the cost of token consumption. System module parameters are fixed; plugin parameters are configurable. A tool description needs to state its purpose and the situations where it should not be called. When a required parameter has no suitable value in context, the model may not call the tool, and a prompt can guide the user to supply the parameter. Models that support function calling can invoke multiple tools at once.
In advanced orchestration, available tools are connected to the tool call module. A connected tool separates tool inputs from ordinary inputs, and its description can be edited to make the call timing more precise. Starting with a small number of tools for tuning is advised. The tool call termination node ends the current invocation, calls no further tools, and does not call the AI to summarize tool results. Custom variables extend tool variable inputs so that nodes not treated as tool parameters can still be invoked.
Retrieval results do not constitute authorization. A similarity score does not distinguish business category, entity or permission. Content returned by knowledge base search requires manual checking of subject, object and status, and cannot be treated directly as a basis for approval or execution.
Hypothetical paper exercise sample: a platform campaign operator submits a promotional price approval form on 10 March 2025, with source text "Product A, campaign price 199 yuan, campaign period 15 March to 20 March 2025, approval status is pending review." A separate inventory lock table has source text "Product A, locked quantity 500 units, release condition: campaign end or approval rejection." The proposed sentence to be judged is "Product A has been approved to join the 15 March campaign at 199 yuan." Checking compares the approval status field in the approval form with "has been approved" in the proposed sentence. Subject and object match, but the approval status is pending review, so the proposed sentence is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Checking method |
|---|---|---|
| Tool description | States "used to check the approval status of promotional price approval forms" | Check whether the description covers input conditions and situations where the tool should not be called |
| Tool parameters | Approval form number, product identifier, campaign period | Check parameter descriptions and required flags, and confirm call behavior when a required parameter is absent |
| Tool connection | Connect the tool call module to the approval form checking tool | Check whether tool inputs and ordinary inputs are separated after connection |
| Tool call termination | Placed after the approval form checking tool | Confirm that after termination no other tool or AI summary is called |
| Custom variable | Add a variable description for the inventory lock table node | Check whether the variable description allows that node to be invoked |
| Prompt guidance | Guide the user to supply the approval form number | Observe whether the model requests the missing required parameter |
Three common errors
- Treating pending review as approved. In the hypothetical sample the approval form source text reads "approval status is pending review," while the proposed sentence says "has been approved." The input is the approval form source text and the proposed sentence; the comparison is between the approval status field and the wording "has been approved." Pending review does not constitute approval, so the proposed sentence is not supported.
- Treating locked inventory as sellable inventory. In the hypothetical sample the lock table source text reads "locked quantity 500 units, release condition: campaign end or approval rejection." The input is the lock table source text and the proposed sentence; the comparison is between locked quantity and sellable quantity. A locked quantity does not indicate sellable stock, so a proposed sentence reading "500 units sellable" is not supported.
- Treating a retrieval result as authorized execution. In the hypothetical sample a knowledge base search returns one approval form record. The input is the retrieval result and the proposed sentence; the comparison is between the retrieved content and authorization status. A retrieval return does not constitute approval, so a proposed sentence reading "authorized for execution" is not supported.
How to verify the setup
- Use the hypothetical approval form source text "approval status is pending review" and the proposed sentence "Product A has been approved to join the 15 March campaign at 199 yuan" as input. Compare the approval status field with the wording "has been approved." The expected result is that the proposed sentence is not supported. If the proposed sentence is changed to "approval status is pending review," it matches the source text.
- Use the hypothetical lock table source text "locked quantity 500 units, release condition: campaign end or approval rejection" and the proposed sentence "Product A has 500 units sellable" as input. Compare locked quantity with sellable quantity. The expected result is that the proposed sentence is not supported. If the proposed sentence is changed to "locked quantity 500 units," it matches the source text.
- Use the hypothetical knowledge base search return record and the proposed sentence "Product A is authorized to execute the promotional price" as input. Compare the retrieved content with authorization status. The expected result is that the proposed sentence is not supported. If the retrieval return is empty, downstream design can be checked against the paper input []; no claim is made that a hypothetical rule search necessarily returns [].