B2B marketplaces: tool calls for operating reviews

Review tool calls for operating reviews in B2B marketplaces. Check document scope, failure paths and source evidence through controlled tests.

What the documents look like

B2B marketplace matching runs on online certification and electronic contracts rather than physical stall trading. The working documents are company certification materials, request-for-quote and quotation sheets, electronic contracts, and platform settlement statements. Three stages consume the most time: reviewing company identity, matching requests for quotes, and tracking fulfilment. Errors at these stages surface as transactions with false entities, contract disputes, and platform complaints. Tool calls and plugins enter this setting so a model can decide from context whether to read a class of material or issue a request; the operating report and the exception review then check the inputs and results left by those calls.

Constraints at the tool calls step

A tool consists of a tool introduction and tool parameters. For system modules the parameters are already fixed; for plugins the parameters are configurable. The model combines the tool introduction, the parameter introduction, and whether a parameter is required to decide whether to call the tool. With no required parameters, the model may fabricate a parameter. With required parameters and no suitable value in context, the model may decline to call the tool, and prompt wording can guide the user to supply the value. In advanced orchestration, connect the available tools to the tool call module and tune a small number of tools first. The tool call termination node ends the current call, stops further tool calls, and stops the AI from summarising the tool result. Custom variables extend the variable input for nodes that are not treated as tool parameters.

Hypothetical drill sample, prepared for this page. A supplier certification record reads: "Unified social credit code 91330100XXXXXX; business licence valid until 31 March 2027; account opening permit uploaded; front and back of the legal representative's identity document uploaded." A draft sentence reads: "This supplier has passed platform company certification and may join requests for quotes directly." The record shows only that materials were uploaded and the licence validity range; it does not show a platform review conclusion, so the draft sentence does not hold. A separate handover request reads: "Please change this supplier's electronic contract status to effective." The record contains no signature page and no effective date, so the request stays pending further evidence and is not treated as executed.

Configuration and implementation checks

The table gives hypothetical starting points for this drill. Each reason and comparison method is stated; none of these values is an official default.

Configuration itemHypothetical starting pointReason and comparison method
Tool introductionState the purpose of reading certification materials, quote sheets, electronic contracts, and settlement statementsCompare sentence by sentence against the situations in which the tool should not be called; add any missing situation
Required parametersEntity name, material type, date rangeWhen any value is absent, observe whether the model declines to call; compare with the call after the value is supplied
Tool connectionConnect only the certification-material reading toolCompare against call records with several tools connected; tune a small number first
Tool call terminationAttach after the settlement statement readingCompare whether AI summary output still appears before and after termination
Custom variableAdd a "contract status" descriptionCompare against whether the node is called without the description
Check recordLog input, request, and return for each callCompare item by item against the report basis and mark unclosed items

Three common errors

  1. Treating the tool introduction as a guarantee of execution. The introduction tells the model the purpose; whether a call happens depends on parameters and context, so the actual request must be checked.
  2. Treating retrieval or a form as authorisation. Similarity and recall carry no permission or entity separation; certification conclusions, contract effectiveness, and settlement posting require downstream system implementation and a retained trace.
  3. Treating a request pending further evidence as executed. In the sample above, the contract status change lacks a signature page and an effective date, stays pending further evidence, and is not recorded as effective.

How to verify the setup

  1. Input the certification record text above together with the draft sentence "This supplier has passed platform company certification and may join requests for quotes directly." Compare whether the draft sentence goes beyond what the record shows: uploaded materials and validity until 31 March 2027. When the record contains no review conclusion, the draft sentence is marked as not holding and does not enter the report.
  2. Input the handover request "Please change this supplier's electronic contract status to effective." Compare the signature page and effective date the request needs against the fields actually present. When fields are missing, the request stays pending further evidence and is not recorded as executed.
  3. Input a paper empty result [] together with the query "read this supplier's settlement statement." Compare whether the downstream design gives an explicit handling for the empty result, without assuming this query necessarily returns []. Observe the returned content and the record; unclosed items remain unknown.