What data for this category looks like
Daily documentation for general mechanical equipment assembly includes assembly process cards, multi-level BOMs, torque records, gap adjustment records, commissioning reports, factory inspection sheets, and packing lists. Data mainly comes from PLM/ERP systems, part inspection data, assembly tool collection terminals, test benches, MES systems, and customer technical agreements. Documents are updated per device and per order. They include numerical fields with units, such as torque in N·m and gap in mm. Some documents have hierarchical nested structures, such as BOMs that include sub-part lists and assembly sequence instructions.
Constraints imposed by these characteristics on tool calling and plugins
Documents are updated per individual device and per order. Tool calling must support dynamic fetching of the latest data. Static knowledge base caching cannot be used. Multi-level BOMs and scattered assembly records require cross-document cross-checking. Tools must handle structured and unstructured data in different formats. Numerical fields with units need automatic parsing and verification by plugins. This prevents errors from unit mismatches. Real-time commissioning data collected by test benches must be compared in real time with offline documents. This requires plugins to support low-latency external interface interactions during calls.
The following configurations are developed for FastGPT v4.17.0.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Workflow/API tool selection | auto | Allows the model to independently determine whether to call plugins to handle different types of document cross-review tasks |
| Workflow/API function-call configuration | allowed | Enables custom plugin calling permissions, to adapt to cross-data-source verification requirements for multiple document types |
| Plugin call timeout | 600 seconds | Adapts to the delay in test bench data return, to avoid interrupting the verification process due to timeout |
| Number of results per single plugin call | Top 3 | Filters redundant data, focusing on core component matching and assembly parameter verification |
| File upload trigger rule | Triggered for assembly-related documents | Automatically enables tool calling mode only when uploaded documents include assembly process cards or BOMs |
| Custom plugin associated scenario | Assembly document cross-review | Limits the scope of plugin calls, to avoid unnecessary calls consuming resources |
Three Common Misconfigurations
- Phenomenon: Custom plugins are not triggered after configuration, and no tool call records appear in the interface. Cause: The plugin is not associated with the "assembly document cross-review" scenario, or
toolChoiceis set tonone, which restricts the model's tool calling permissions. - Phenomenon: The model does not read uploaded assembly documents, and only returns generic responses. Cause: The file upload trigger rule for tool calling is not configured correctly. The model fails to identify the need to read document content.
- Phenomenon: A timeout error returns when calling an external test bench data interface, with a 504 status code. Cause: The plugin call timeout is set too short, and does not adapt to the actual delay of test bench data return.
How to Verify the Configuration
- Upload the BOM document and corresponding assembly process card for a single device, and submit a query asking to "verify component consistency between the BOM and assembly process card". Check if corresponding plugin call entries appear in the interface's tool call records.
- After plugin call is triggered, check if returned results include component matching tolerance data obtained from external interfaces. Confirm the plugin is called normally and returns valid content.
- Upload only non-assembly documents containing plain text, and confirm the model does not trigger tool calls. Verify the effectiveness of the trigger rule.
- Simulate a timeout scenario: set the plugin call timeout to a value lower than the actual delay, and submit a query that requires calling test bench data. Confirm the interface returns a timeout prompt, to verify the timeout configuration is active.
The values provided are common starting points and should be measured against the reader's own samples.