Data covered by this category
Data for semiconductor packaging and testing includes packaging process sheets, bonding and molding process parameters, wafer incoming inspection maps, test program versions, yield reports, failure analysis reports, and more. Data sources include wafer test maps, MES production systems, bonding and molding equipment, ATE test machines, X-ray/SAM inspection equipment, and customer specification requirements. Update frequency is per production batch, after each individual chip test, and daily updates. Documents include fixed fields such as batch number, process thresholds, test bin values, and failure types. Units cover industrial standard units including voltage, current, time, and temperature.
Constraints imposed on model access and configuration
Since this category’s data sources cover real-time collection from multiple devices and high-frequency updates, model access must support real-time synchronization verification for multi-source data. Fields vary significantly across different documents, so adaptive context clipping rules must be configured to prevent truncation of key process parameters. Fine-grained updates per batch and per chip require the model’s context length to be adapted for small-batch data processing, avoiding redundant loading. Failure analysis reports and yield reports have differing structural levels, so parameters must be adjusted to support recall and Q&A logic for non-standard text. Additionally, some packaging and testing factories require intranet deployment, so model access must support private address configuration to ensure data security.
Configuration settings
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
maxContext | 8000–12000 token | Adapts to the context length of multi-document concatenation such as packaging process sheets, test program versions, and yield reports, avoiding truncation of key process parameters |
maxTokens | 2000–3000 token | Matches the output length requirement for a single process Q&A, ensuring complete coverage of parameter descriptions and operating steps |
quoteMaxToken | 1500 token | Limits the length of a single recalled document segment, adapting to the structural fragment length of wafer incoming inspection maps and failure analysis reports |
temperature | 0.1–0.3 | Reduces random generation probability, ensuring accuracy of process parameters and standard Q&A, avoiding outputs that do not comply with industry specifications |
vision | Enabled | Adapts to visual Q&A requirements for wafer test maps and X-ray inspection images, helping align physical batches with drawing data |
AIPROXY_API_ENDPOINT | Fill in according to the deployment environment | Adapts to access addresses for local or cloud-based large models, meeting security requirements for intranet deployment in packaging and testing factories |
Three common configuration errors
- A version incompatibility error occurs when calling a local large model. This is caused by failing to install the dependency library specified by the model version, and using the official latest version directly.
- Interface connection failure occurs when configuring the model access address. This is caused by misuse of the old OneAPI service address, and not using the proxy address format supported by FastGPT v4.17.0.
- Recalled process document segments are too long, causing context overflow. This is caused by failing to set a reasonable
quoteMaxTokenparameter, and not limiting the length of a single recalled document segment.
How to verify successful configuration
- Upload a single packaging process sheet, initiate a process parameter query, and check whether the returned result includes core process parameters such as accurate bonding temperature and molding pressure.
- Upload a wafer test map, initiate visual Q&A, and check whether the model can correctly identify test bin values and corresponding failure types.
- Submit yield reports for multiple batches, initiate a batch alignment query, and check whether the model can correctly associate batch numbers with corresponding test data.
- View the system call log, confirm that the
maxContextparameter for large model calls has been correctly passed, and no context truncation errors are present.
The values provided are common starting points and should be measured against the reader's own samples.