What the data for this category looks like
The business data for photovoltaic module encapsulation originates from the cell grading system, glass adhesive film backplane batch ledger, production MES system, string welding and lamination equipment, EL tester, and IV tester. One record is created for each module and each material batch, with data updated accordingly. Daily documents include layout drawings, material batch records, lamination process records, EL crack images, IV test reports, insulation withstand voltage records, and packing lists. Data fields include cell current gear, lamination temperature and time, EL crack location, IV power value, and more. Units include degrees Celsius, seconds, watts, batch numbers, and other standard units.
What constraints these characteristics impose on the "tool calling and plugins" workflow
Photovoltaic encapsulation has high real-time requirements. Batch data for each module is updated continuously, so tool calling must support real-time pulling of the latest production and test data. Business documents use multiple formats: images (EL images), tables (IV reports), and text (process records). Multiple plugin types must be configured to adapt to parsing and pulling of these different formats. Data sources are scattered across multiple devices and systems, so multiple independent HTTP plugins are required to connect and retrieve data. The accuracy of Q&A results directly impacts production compliance and product quality. Tool calling must strictly align with process specifications and technical standards to prevent production deviations caused by incorrect calls.
How to configure the settings
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Workflow/API tool selection | auto | Adapt to the multi-tool calling scenario of photovoltaic encapsulation, allowing the large model to automatically match the tools required for process query, EL interpretation, and power verification |
| Workflow/API function-call configuration | Force trigger | Ensure that process standard Q&A must call tools to obtain real-time batch data, avoiding reliance on expired cached data |
| HTTP plugin configuration | Bind 3 dedicated data source plugins: MES batch records, EL tester data, IV test reports | Cover the real-time pulling requirements of core photovoltaic encapsulation business documents |
| Single tool call limit | Top 2 results | Match the single batch data volume per module, avoiding redundant information interfering with Q&A accuracy |
| Timeout threshold | 600 seconds | Adapt to time-consuming scenarios such as EL image parsing and MES data pulling, calibrated based on actual measurements |
| Plugin file size limit | 100 MB | Adapt to the conventional size range of documents such as EL images and IV test reports |
Three common configuration errors
- Phenomenon: The published API returns the raw output of tool calls without using it as context input for the large model. Cause: The logic to restrict tool call results to only serve as large model input was not configured, leading to direct output of the raw results.
- Phenomenon: Some model calls display a prompt that tool calling is not supported. Cause: The range of models that support
toolChoicewas not verified, and an incompatible model was used. - Phenomenon: HTTP plugins fail to pull MES or EL tester data. Cause: The plugin’s API endpoint and authentication configuration were not completed correctly, resulting in failed requests.
How to verify a successful configuration
- Trigger a Q&A related to process specifications, and check if the corresponding tools are automatically called to pull the latest lamination process records and cell batch data.
- Call the published API interface, and confirm that the returned results only include answers organized by the large model, with no raw tool call data displayed directly.
- After importing the HTTP plugin, manually trigger a test to confirm that MES batch records and EL image analysis results can be pulled normally.
- Swap to different models that support
toolChoice, and verify that the tool calling function operates correctly.
The values provided are common starting points and should be measured against the reader's own samples.