What this category’s data includes
Core data for this category originates from process and quality departments of foundries and die casting workshops. Daily documents include casting process cards, batching lists, heat logs, mold history records, first-piece inspection records, material test reports, and nonconforming product review forms. Data sources cover ERP material master data, melting furnaces and thermometers, spectrometers, MES heat logs, mold management systems, customer drawings, and technical requirements. Updates run per heat, per batch, or daily. Document fields include heat number, alloy batching ratio, melting temperature, mold cycle, defect type, and more.
Constraints for forms and interactions
Since data updates per heat, per batch, or daily, forms must link unique identifier fields such as heat number and mold cycle. This ensures Q&A content can be traced to specific production units. Frequent tasks like checking batching ratios and interpreting parameters require forms to support preset high-frequency options. This reduces manual input errors. Multiple structured document types require forms to associate fields from specific documents for quick parameter retrieval. The core requirement of defect traceability mandates that forms include mandatory fields such as defect type and problem description. This prevents unsubstantiated Q&A submissions.
Configuration settings
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
formInput | Bind three mandatory fields: heat number, mold cycle, and defect type. Limit field length to under 100 characters | Heat number and mold cycle are core traceability identifiers. Ensure complete input to avoid unsubstantiated Q&A |
userSelect | Preset two drop-down options: alloy batching ratio and common defect types. Limit option count to no more than 20 | Reduce manual input errors to support fast operations in frequent inspection and maintenance scenarios |
| Q&A pair extraction trigger timing | Split user input first, then apply Q&A split guide words | Splitting long text into paragraphs before extracting Q&A pairs aligns better with the structured nature of casting parameters |
| User input verification threshold | 800–1200 character range, calibrated via actual testing | Casting parameter descriptions are mostly medium-to-long text. Excessively long text increases model processing load |
| Context call limit | Top 5 historical interaction records | Equipment inspection and maintenance context typically relates to a single heat or mold cycle. Excess context can interfere with current question processing |
| Model output format control | Force structured results with two fields: defect cause and maintenance suggestion | Streamline organization of inspection and maintenance records for quick on-site workshop reference |
Three common configuration errors
- Workflows fail to trigger form validation, allowing users to submit questions without entering a heat number. This occurs when mandatory field validation rules for
formInputare not configured. - The qwen-plus model does not output a thought process. This happens when the thought output switch is not enabled in model settings, or the prompt does not include thought step guidance.
- Extracted Q&A pair counts do not match expectations, with only partial content retrieved. This occurs when Q&A split guide words are applied without first splitting the input content, preventing full extraction of long text.
How to verify correct configuration
- Submit the form after filling non-mandatory fields, verify that a missing mandatory fields prompt appears.
- Enter an equipment inspection record longer than 1200 characters, verify that an input length limit prompt triggers.
- Select a preset defect type option, verify that the drop-down menu loads normally.
- Call the model to process a question with a heat number, verify that the returned result links traceability information for the corresponding heat.
The values provided are common starting points and should be measured against the reader's own samples.