What the data for this category looks like
Data for this category originates from cell grading ledgers, glass adhesive film backplane batch records, workshop MES systems, string welding and lamination equipment logs, and EL/IV tester output files. Data is updated with one record per module and per material batch. Documents involved in daily audits include layout drawings, material batch sheets, lamination process parameter sheets, EL crack images, IV power test reports, insulation voltage resistance records, and packing lists. Data fields include cell current gear values, lamination temperature and time, EL crack coordinates, IV test power values, insulation resistance values, and others. Some data is structured tables, some are images and binary files, with no unified fixed format.
What constraints do these characteristics impose on the multi-turn dialogue and prompt engineering link
The data characteristics of this category impose multiple constraints on multi-turn dialogue and prompt configuration. First, data is updated per single module and single batch and includes multiple file types. Dialogue context must accurately bind the variables parameters of the corresponding batch to avoid cross-batch data confusion. Second, the audit process must sequentially complete material matching, process verification, defect judgment, and rectification confirmation. Multi-turn dialogue must follow the process progression, and prompts must clearly define the task scope for each round to prevent context overload. Third, some data is unstructured images. Prompts must specify the priority and parsing rules for image retrieval to ensure the accuracy of EL crack judgment. Finally, audit errors directly relate to customer claims. The dialogue must retain a complete audit trail for retrospective verification.
How to set configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Maximum context length | 8000–12000 characters | Single-batch audit data for photovoltaic modules includes multiple documents. Sufficient context is required while avoiding response overload |
variables binding fields | Cell batch number, lamination process parameters, EL image ID | Accurately match the exclusive data of the currently audited module to avoid retrieving irrelevant batch records |
| Single dialogue round limit | 12 rounds | The audit process usually consists of four stages: material verification, process verification, defect retest, and rectification confirmation. 12 rounds can cover the complete audit cycle |
messages context truncation threshold | Retain the most recent 6 rounds of dialogue | Filter redundant historical information and focus on issues and responses from the current audit stage |
chatId session isolation rule | Bind by supplier batch + component serial number | Ensure that session data does not interfere with each other when auditing multiple batches simultaneously |
| Question optimization switch | Enabled | Automatically convert colloquial questions from auditors into structured retrieval instructions to adapt to the retrieval needs of multiple types of documents |
Three common configuration mistakes
- Phenomenon: Dialogue output response time exceeds the preset threshold. Cause: Maximum context length is not restricted, and too much large file data such as EL images and IV test reports is loaded during each retrieval
- Phenomenon: File transfer fails when calling the dialogue API. Cause: File paths are not pre-bound in
variablesor the file upload whitelist is not configured - Phenomenon: Images are not fully displayed during MCP calls. Cause: The prompt does not specify image resolution adaptation parameters, or the batch data corresponding to the EL image is not associated
How to verify successful configuration
This verification process is based on FastGPT v4.17.0.
- Test a single-batch module audit dialogue. Check that the context only retains the
variablesbound data of the current batch - Initiate a multi-turn dialogue. Verify that the output response time meets the preset threshold
- Upload EL images and IV test reports. Confirm that the dialogue can correctly associate corresponding fields and generate audit conclusions
- Initiate multiple concurrent sessions. Check that no session conflict errors occur
The values provided are common starting points and should be measured against the reader's own samples.