What the data for this category looks like
Core documents for this category include packaging process cards, bonding and molding parameters, wafer incoming maps, test program versions, yield reports, and failure analysis reports. Data sources include wafer test maps, manufacturing execution systems (MES), bonders, molding equipment, ATE test machines, X-ray/SAM inspection equipment, and customer specification documents. Update frequency for data aligns with production batches, single-chip testing, and daily operations. Fields across different documents carry exclusive identifiers such as wafer numbers, bonding temperature parameters, test bin codes, and yield values.
What constraints these characteristics impose on multi-turn dialogue and prompt setup
This category’s multi-source heterogeneous documents, high-frequency update cadence, and fine-grained field traits create clear constraints for multi-turn dialogue and prompt configuration. Documents in varying formats need explicit matching rules for their corresponding fields in prompts to prevent cross-document field confusion. High-frequency updated production data requires limiting the valid scope of historical messages in dialogue context, to stop expired data from disrupting current reviews. Test data tied to individual chips and batch-bound attributes must link exclusive identifiers via variables, so multi-turn dialogue can accurately trace document content for the target batch. Data from cross-device sources needs explicit field associations for corresponding systems in prompts, to avoid matching errors between ATE test data and MES batch data.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
messages | Retain the most recent 20 conversation messages, truncate excess content | Production data for this category does not require long historical context to avoid reasoning overload |
variables | Bind two core variables: batch number and wafer ID | All documents for this category are tied to exclusive production identifiers, requiring variables to accurately match corresponding data |
chatId | Generate a unique chatId per production batch | Document data for different batches is isolated, requiring separate dialogue context for each batch |
| Maximum context length | 8000–12000 characters | Total document length per batch for this category is moderate; excessive length causes reasoning delays |
| Question optimization | Enabled | This category contains a large number of technical terms; optimization improves precise matching of prompt content |
| Single historical retention duration | 24 hours | Production data for this category updates daily; expired context does not need to be retained |
Three common misconfigurations
- Phenomenon: Dialogue history shows wafer incoming map content from a non-current batch, leading to batch matching errors in review results. Cause:
chatIdis not bound per production batch, and dialogue contexts for different batches are not isolated. - Phenomenon: The test bin code output by the prompt does not match the wafer ID, making association with the corresponding yield report impossible. Cause: Exclusive production identifiers are not bound via
variables, and the prompt does not specify clear field association rules. - Phenomenon: The dialogue fails to return results within 30 seconds, or returns empty review results. Cause: The number of historical
messagesis not limited, leading to reasoning timeouts due to overly long context.
How to confirm the configuration is correct
This verification process uses FastGPT v4.17.0. Complete these steps:
- View the
chatIdfield of the conversation to confirm each production batch has a unique identifier. - Check whether
variablesbind exclusive production fields such as batch number and wafer ID. - Test multi-turn dialogue to confirm only the most recent 20 historical messages are retained, with no expired data remaining.
- Run a single-batch review workflow to confirm the prompt correctly matches fields from the corresponding documents.
Note: The configuration values listed here are common starting points. Adjust them based on your own specific samples and use cases.