What the documents look like
Brand outlet discount operations exist to clear out-of-season stock. The category does not maintain a full-price new-arrival image, so the working materials are discount approvals, slow-moving inventory lists, price tags and gross-margin worksheets. Data updates weekly or monthly. A comparison between contract and transaction materials has to place approval records, list entries, price-tag information and worksheet figures under the same entity, the same period and the same field before any conclusion is drawn.
Tool calls and plugins enter this step because the model decides from context whether to invoke a comparison, retrieval or calculation tool, and the tool result is returned to the model for summarisation. A tool description states the purpose of the tool and the input conditions; it also states situations in which the tool should not be invoked. Tool parameters divide into fixed parameters for system modules and configurable parameters for plugins. Where a required parameter has no suitable value in context, the model may decline to call the tool. Where a tool has no required parameters, the model may still fabricate a value. Retrieval on its own does not constitute authorisation, and a similarity score does not perform permission or entity isolation.
Hypothetical paper exercise. On 10 March 2025 a brand outlet channel clearance group submits a discount approval form. The form reads: "Request to clear out-of-season coat style A-1023 at a 40% discount, valid 11 March 2025 to 31 March 2025, approval status is pending review." On the same day a gross-margin worksheet states: "Style A-1023, tag price 899 yuan, 40%-off selling price 539.4 yuan, estimated gross margin -3%." A draft sentence reads: "This discount request has been approved and the 40% clearance may proceed." The request to be judged reads: "Compare the approval form with the worksheet and confirm whether the discount may be executed."
Constraints at the tool calls step
Tool invocation depends on the model reading context and deciding whether to execute. A tool with no parameters is judged from its description alone. For a tool with parameters and no required parameter, the model may call it even without a suitable value, and may fabricate one. For a tool with a required parameter and no suitable value in context, the model may not call it; a prompt can guide the user to supply the value. In advanced orchestration, available tools are connected to the tool call module, and a small number of tools is tuned before more are added.
The tool call termination node ends the current tool call. It does not call other tools and does not call the AI to summarise the tool result. This matters where a direct comparison output is wanted rather than a summarised answer.
The paper exercise carries several evidence gaps that must remain gaps. The approval form records "pending review"; a pending request is not an approval, and no signature, effective date or authorisation clause may be supplied to close the question. The worksheet records an estimated margin of -3%, which is a calculated figure and not an authorisation to execute. The validity window 11 March 2025 to 31 March 2025 belongs to the discount execution period; it is not a supply period and the two fields may not substitute for each other. The worksheet does not state an entity, so the entity match between the two documents remains unverified. Retrieval results, if any, sit inside the authorised target scope or an explicit manual input; where that scope is missing, the gap is recorded rather than filled.
Configuration and implementation checks
| Configuration item | Setting or check for this page | Relation to the sample |
|---|---|---|
| Tool description | State the purpose of comparing an approval form with a worksheet; list input conditions and situations where the tool should not be invoked | The model uses this to decide whether to call the comparison tool |
| Tool parameters | Check plugin configurable parameters so entity, style number and period fields can be passed in | The sample needs style A-1023 and the approval period passed in |
| Required parameters | Where no suitable required parameter exists the model may not call the tool; a prompt guides completion | A missing approval-status field is recorded as a gap |
| Tool connection | In advanced orchestration, connect tools to the tool call module and tune a small number first | Comparison and retrieval tools are connected separately |
| Tool call termination | Placed after the comparison tool to end this call without AI summarisation | Used when a direct comparison result is required |
| Hypothetical trial starting value | For this paper exercise a similarity reference value of 0.75 is set, because a text pair is being compared and the comparison object is the score of different text blocks under one query | This value is not an official default; it serves only this paper check |
Three common errors
- Treating "pending review" as approved. The input is the approval form text; the field compared is approval status; the observable result is that "pending review" and "approved" are mutually exclusive values on the same field, so the draft sentence is unsupported. If the form had not stated a status, the item would be recorded as awaiting verification rather than as a conflict.
- Treating a retrieval score as a permission filter. The input is a similarity score returned by one retrieval; the object compared is the authorised-scope list; the observable result is that a high or low score does not change whether a record falls inside the authorised scope. Where the authorised list is absent, the gap is recorded.
- Treating the validity window as a supply period. The input is the approval form text; the fields compared are validity and supply period; the observable result is that these are different fields and one does not replace the other.
How to verify the setup
- Input the sample approval form text and the draft sentence; compare the value of the approval-status field. The expected observation is that "pending review" and "approved" are mutually exclusive, so the draft sentence is judged unsupported. If the form does not state an approval status, record the item as awaiting verification.
- Input the sample worksheet text and the request to be judged; compare the style number, selling price and gross-margin fields. The expected observation is that style A-1023 matches the approval form, that the -3% margin is a calculated figure, and that it does not constitute an authorisation conclusion for executing the discount. If the worksheet does not state a style number, record an evidence gap.
- Input a paper retrieval result consisting of an empty array; compare whether the downstream tool is designed to accept an empty result. The expected observation is that the downstream step does not produce a fabricated record. This check does not assert that a hypothetical rule-based retrieval necessarily returns an empty array; it uses a clearly stated paper input to examine the downstream design. Libraries built on different vector models cannot be associated together; identifiers are checked first, non-matching items are excluded, and the actual observation is recorded.