What data looks like for this category
Data for general mechanical equipment assembly scenarios typically comes from PLM/ERP, parts inspection systems, assembly tool collection terminals, test benches, MES, and customer technical agreements. Daily documents include assembly process cards, multi-level BOMs, torque records, clearance adjustment records, commissioning reports, factory inspection certificates, and packing lists. Fields include process number, operator, measured torque (unit: N·m), clearance value (unit: mm), coaxiality tolerance (unit: μm), and others. Data is updated per single equipment or single order, with new associated datasets generated once one equipment is assembled or an order is delivered.
What constraints do these characteristics impose on the tool calling and plugins link
Since data is updated per single equipment or single order, tool calling must bind the context of the corresponding instance to avoid mixing cross-order data. This category assembles multiple varieties per order, relies on component fitting and debugging, and differs from dimensional qualification judgment for single-part machining. Therefore, tool calling must support multi-level BOM associated retrieval and multi-source process data integration. Process parameters with units must have their field units accurately matched by tool calling, otherwise analysis result deviations will occur. Additionally, closed-loop fault analysis for assembly tests requires associating data from multiple systems, so tool calling must support pulling data across data sources.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Workflow/API tool selection | auto | Automatically select tools or knowledge base retrieval for different quality traceability scenarios |
| Workflow/API function-call configuration | enabled | Support calling assembly-specific tools such as BOM query and torque data extraction |
| Knowledge base binding rules | Match by order number/device serial number variable | Data is updated per single equipment or single order, requiring binding of the corresponding instance context |
| Number of items returned per tool call | Top 8 | Assembly documents have multiple levels, control return volume to avoid context overflow |
| Tool call timeout | 600 seconds | Closed-loop analysis of assembly tests requires retrieving multiple types of log data, which takes a long time |
| Context window limit | 8000 token | Adapt to the context length after combining multiple documents in assembly scenarios; FastGPT v4.17.0 default upper limit matches this requirement |
| Variable mapping verification | Calibrate based on actual measurements | Field names and units vary across different devices, requiring alignment with actual business needs |
Three common misconfigurations
- Phenomenon: Tool call return results are unrelated to the current order, or redundant cross-order data is retrieved. Cause: The corresponding knowledge base is not bound by order number/device serial number, and variable matching rules are not correctly configured.
- Phenomenon: Tool call nodes in advanced orchestration fail to execute, and external tools cannot be triggered. Cause:
functionCallis not set toenabled, and tool calling function is not enabled. - Phenomenon: The request body of each AI call cannot be viewed, or debug logs are empty. Cause: FastGPT debug log switch is not enabled, or request log recording is not enabled in system settings.
How to confirm configurations are correctly set
- Enter a specified order number and device serial number to trigger a quality traceability query, and verify that the returned knowledge base results only include assembly documents for that order.
- Call the BOM query tool, enter a specified material code, and verify that the returned fields include parameters such as torque and clearance required by assembly process specifications.
- Trigger a tool call node, view the system execution log, and confirm that the configurations of
toolChoiceandfunctionCallhave taken effect. - Call the local model plugin, enter a standardized test prompt, and verify that the returned results meet expectations with no timeout errors.
The values provided are common starting points and should be measured against the reader's own samples.