What the documents look like
In the foreign exchange receipt, settlement and trade directory category, contract and transaction material comparison centres on four document types: foreign exchange receipt advices, trade contracts, invoices and foreign exchange business registration materials. The category differs from letter of credit document examination: the focus falls on the authenticity of foreign exchange funds and the underlying trade background, and a full set of letter of credit documents is not examined. When comparison goes wrong, the direct consequences include bank rejection of documents, abnormal foreign exchange classification and funds that cannot be credited to the account. Data updates occur per transaction and per month, so tool calls need the ability to distinguish per-transaction verification from monthly aggregation.
For tool calls and plug-ins, the practical constraints follow from how the tool call node decides. The model decides whether to call a tool on the basis of three pieces of information: the tool description, the parameter description and whether a parameter is required. System module parameters are fixed; plug-in parameters are configurable. When a required parameter has no suitable value in context, the model may not call the tool, so the missing value must be supplied through interaction and the actual request checked. In advanced orchestration, connecting a tool to the tool call module automatically separates tool inputs from ordinary inputs, and the description can be edited to make the call timing more precise. Starting with a small number of tools and tuning before adding more is the documented approach. A tool call termination node attached after a tool forcibly ends the current tool call: no further tools are called and the AI is not called to summarise the tool result. Custom tool variables can add parameter descriptions for nodes that are not treated as tool parameters.
A hypothetical paper exercise sample is set out here for illustration only. The exporter name below is an English rendering of the original Chinese company name, not a verified registered English name. The material text reads: "Exporter: Huayue Light Industry Manufacturing Co., Ltd.; Contract No.: HY-2024-0312; Contract amount: USD 186,000; the receipt advice shows payer: HWAYUE LIGHT IND.; receipt amount: USD 186,000; receipt date: 28 March 2024; Invoice No.: INV-0312; invoice amount: USD 186,000." The draft sentence to be assessed reads: "The payer named on the receipt advice is the same entity as the contract exporter, the amounts agree, and settlement registration can proceed." In this sample, no mapping is supplied between payer HWAYUE LIGHT IND. and the exporter identified by the original Chinese name, rendered above in English. The receipt advice also lacks a field directly linking it to the contract number. Matching amounts establish equality of the amount field only, not entity identity. The conditions and supporting documents for settlement registration are not supplied and remain pending verification.
Constraints at the tool calls step
The tool call node operates on the tool description, the parameter description and the required flag. A tool with no parameters is judged solely on its description. A tool with parameters but no required parameters can still be called when no suitable value exists in context, and the model may fabricate a value. A tool with required parameters may not be called when no suitable value exists, and the prompt can guide the user to supply the value. In models that support function calling, multiple tools can be called in one pass.
The termination node ends the current call and stops both further tool calls and AI summarisation of tool results. Custom tool variables extend variable inputs for nodes not treated as tool parameters, and the tool call then invokes that node and the workflow after it. Retrieval and recall do not perform permission or entity isolation. An empty result can be checked downstream against a paper input of []; it cannot be asserted that a fictional rule search returns [] under all conditions. Libraries using different vector models cannot be associated together; identifiers must be checked first and non-matching libraries excluded. Optional paths in a documented flow diagram do not mean every request executes all of them: image retrieval requires an image input and model capability, and reranking requires a model and must be enabled in settings.
Configuration and implementation checks
| Configuration item | Setting content | Verification method |
|---|---|---|
| Tool description | State input conditions and situations where the tool should not be called | Check the description text against the four document types on this page |
| Parameter description and required flag | Mark contract number, amount and entity name as required or not | Observe actual calls and trace sent parameter values when a required parameter is missing |
| Tool connection | Connect the tool to the tool call module in advanced orchestration | Confirm that tool inputs and ordinary inputs are separated after connection |
| Tool call termination | Attach after the comparison tool to end the current call | Confirm that no further tools and no AI summary are invoked |
| Custom tool variables | Add parameter descriptions for nodes without parameters | Check that the parameter description corresponds to the actual node |
For this exercise, a hypothetical starting point is set: the contract number is treated as a required parameter, while amount and entity name are treated as non-required parameters. The reason for this choice is that the contract number is the comparison key, and without it the contract record cannot be located. Observe whether a call is actually made without the contract number, and trace every sent parameter to supplied material or a declared variable. A value without provenance is an issue to investigate, not an acceptable completion of the missing input.
Three common errors
- Writing the tool description as a general statement without input conditions or situations where the tool should not be called. The model may then call the tool on unrelated materials. The check is to compare the description against the material list and confirm whether it covers the four document types on this page.
- Treating the parameter description as a guarantee of execution. The parameter description only explains the meaning of the parameter; it does not ensure that the model passes the correct value. Observe whether a call is actually made and trace each sent value to supplied material or configuration. A value without a source remains under investigation and is not an acceptable execution result.
- Expecting AI summary output after the tool call termination node. The termination node ends the current call and does not invoke the AI to summarise the tool result. The check is to confirm the position of the termination node and whether subsequent nodes are skipped.
How to verify the setup
- Verification one: input the hypothetical source text and the revised draft, then compare payer HWAYUE LIGHT IND. with the exporter named in Chinese and rendered above in English. No identity mapping is supplied. Neither the English rendering nor equal amounts establish that they are the same entity, and an unproven mapping is not a contradiction. Keep the identity claim and the unsupported settlement-registration conclusion pending verification.
- Verification two: mark contract number HY-2024-0312 as required. For the paper comparison, supply only receipt-advice fields without a contract number; do not also supply the contract passage containing it. Observe whether a call is actually made and trace each actual request parameter to its source. Record a non-call as observed, and flag any value without provenance for investigation. A fabricated contract number is not a passing result. Without a live test, retain this as a verification plan.
- Verification three: attach a tool call termination node after the comparison tool and input the complete sample materials. Compare the call chain before and after the termination node. The expected phenomenon is that after the termination node executes, no further tools are called and no AI summary is produced. Observe whether the output contains AI-generated summary text.