What the data for this category looks like
Data for this category primarily comes from wafer test maps, MES systems, bonding and molding equipment, ATE test machines, X-ray/SAM inspection devices, and customer specification documents. Daily processed documents include packaging process cards, bonding and molding parameter sheets, incoming wafer maps, test program version files, yield reports, and failure analysis reports. Data updates cover per production batch, single chip test, and daily summary. Fields include batch identifiers, test bin values, bonding pull force parameters, molding temperature, yield data, and more. Multi-source heterogeneous input formats must be supported.
What constraints these characteristics impose on model access and configuration
Multi-source heterogeneous data sources require model access to support different input formats. Multi-modal data parsing must be supported to handle non-text files such as wafer maps and scanned images. High-frequency updated production data requires the context window to adapt to dynamic single-batch data volumes, and avoid truncating critical information. Strict business error consequences require model outputs to remain rigorous, and avoid excessive divergence. Batch identifiers must be linked to corresponding test and failure data, so configuration must ensure the context fully carries all information associated with a batch.
How to set the configuration
This configuration specification is based on FastGPT v4.17.0. Specific configuration items and recommendations are as follows:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
vision | Enabled, to enable multi-modal parsing capabilities | This category requires processing non-text inspection data such as incoming wafer maps and X-ray scanned images |
maxContext | 8000–12000 token | Adapts to the total character count of single-batch packaging process cards, test data, and associated documents |
temperature | 0.1–0.3 | Inspection and audit require strict matching of parameters and document content, and avoid divergent outputs |
AIPROXY_API_ENDPOINT | Fill in according to the deployed local proxy address | Adapts to call requirements for private deployment or local Ollama models |
quoteMaxToken | 2000 token | Ensures complete reference of core document fragments such as test program versions and yield reports |
| Workflow/API tool selection | auto | Automatically call document parsing tools to handle multi-source heterogeneous production data |
Three common errors
- When adding a new model vendor, the Volcano Ark option is not found. The cause is that the third-party vendor access channel was not manually added to the platform's vendor configuration list.
- After configuring Qwen QWEN2.5-VL-7B-instruct, no image analysis results are returned. The cause is that the
visionconfiguration item was not enabled, or no image file in the correct format was uploaded. - A 404 error is returned when adding a local Ollama model via AI proxy. The cause is an incorrectly configured proxy address, or the local model did not start normally.
How to confirm the configuration is complete
- Navigate to the model channel configuration page, and confirm that the target vendor or local proxy address has been added.
- Upload a single wafer test map, trigger multi-modal parsing, and check whether the wafer batch identifier can be extracted.
- Submit a batch of packaging process card data, and check whether the audit results returned by the model include corresponding business fields.
- Adjust the
temperatureparameter and submit a test, and confirm that the rigor of the output changes meets business requirements.
The values provided are common starting points and should be measured against the reader's own samples.