What the Data for This Category Looks Like
Daily documents for this category include packaging process cards, bonding and molding parameters, wafer incoming material maps, test program versions, yield reports, and failure analysis reports. Data sources cover wafer test maps, MES systems, bonding and molding equipment, ATE test machines, X-ray/SAM inspection equipment, and customer specification requirements. Update frequency varies with batch production, single-chip testing, and daily process inspections. Fields include batch identifiers, process temperature and pressure parameters, test bin codes, failure types, and more. Units include degrees Celsius, Pascals, volts, and other standard units.
Constraints on Multi-turn Dialogue and Prompt Engineering
Diverse data sources spanning the entire production process require multi-turn dialogue to gradually align core batch identifiers, preventing cross-batch incorrect data calls. High update frequency requires dialogue contexts to retain the latest process parameters and test results, avoiding references to outdated batch information. Multiple document types including process parameters and test reports require prompt engineering to clearly distinguish field priorities across document types, ensuring the large model accurately matches corresponding data. Severe consequences such as mixed batches and packaging failures require adding verification steps in multi-turn dialogue to confirm aligned information is correct.
Configuration Settings
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Maximum Context Length | 8000–12000 token | Covers the total character count of a single batch's packaging process card, test program version, and yield report |
| Variable Binding | Bind batch numbers and test program versions from variables | Accurately associate exclusive data for the current audit batch to avoid cross-batch incorrect data calls |
| Dialogue History Management | Bind single-batch audit sessions by chatId | Isolate dialogue contexts across different batches to ensure information for each audit round is independent |
| Prompt Template | Fixed to include three stages: batch verification, parameter alignment, failure analysis | Matches the standard process of packaging and testing audits, guiding the large model to output results per business logic |
| Model Call Timeout | 600 seconds | Adapts to the result return cycle of ATE test machines and X-ray scans |
| Knowledge Base Retrieval Count | Top 3–5 entries | Avoids redundant data interfering with the large model's extraction of core audit information |
Three Common Misconfigurations
- Phenomenon: A connection error prompt appears after submitting a third-party model, and the error persists after enabling a proxy. Cause: The API gateway address and secret key for the third-party model were not correctly filled in the FastGPT v4.17.0 model configuration page.
- Phenomenon: The guest login dialogue icon cannot be modified. Cause: A custom icon file that meets format requirements and is no larger than 2 MB was not uploaded in the global dialogue settings.
- Phenomenon: Wafer test maps returned by the knowledge base cannot be displayed in dialogue. Cause: The prompt did not explicitly require converting image links to standard Markdown format, causing the large model to only return plain text links.
How to Verify Successful Configuration
- Initiate a single-batch supplier audit dialogue, check if the
messagesfield of the dialogue includes the current batch's wafer incoming material map number and test program version. - Enter bonding process parameters, verify if the returned results automatically match the corresponding fields in the batch's packaging process card.
- Trigger a model connection test, confirm no connection error-related errors are displayed.
- Enter questions related to failure analysis, verify that returned results are associated with the batch's failure analysis report data.
The values provided here are common starting points and should be evaluated against the reader's own samples.