This document was written based on FastGPT v4.17.0.
What Data for This Category Looks Like
Data for this category covers records related to the entire semiconductor packaging and testing process. Daily documents include packaging process cards, bonding and molding parameter tables, incoming wafer maps, test program version files, yield reports, and failure analysis reports. Data sources include wafer test maps, manufacturing execution systems (MES), bonding and molding equipment logs, ATE tester outputs, X-ray/SAM inspection results, and customer specifications. Data is updated synchronously by production batch, single-chip test results, and daily production progress. Fields include batch number, wafer ID, test bin value, yield value, equipment parameter thresholds, and other related items.
What Constraints These Characteristics Impose on the Multi-turn Dialogue and Prompt Engineering Link
Configuration for multi-turn dialogue and prompt engineering for this category must match the high-frequency update and multi-source characteristics of the data. First, multi-turn context must be limited to current day’s or current batch’s data to avoid mixing parameters across batches. Second, prompts must clearly specify the units and sources of data fields. For example, distinguish between ATE test bin values and X-ray inspection failure types. Third, since data comes from multiple types of equipment and systems, prompts must require the model to prioritize calling the latest synchronized production data. Additionally, limit the number of documents that can be processed in a single dialogue to prevent exceeding context capacity. Finally, ensure that batch numbers and wafer IDs mentioned in dialogue strictly correspond, to avoid cross-batch parameter reference errors that reduce the accuracy of bid documents.
How to Configure the Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chatId | Bind to the current bid batch ID | Ensure multi-turn dialogue only links to production data for the corresponding batch, avoiding cross-batch confusion |
messages | Retain the first 6 rounds of dialogue content | Match the communication context scope for a single batch bid, covering core interactive information |
variables | Preset batch number, wafer ID, test bin value | Quickly call core business fields for packaging and testing, reducing repeated input |
| Maximum context length | 8000–12000 characters | Cover the total character count of packaging documents for a single batch, preventing key parameters from being truncated |
| Session timeout period | 300 seconds | Adapt to the interaction duration for a single bid document preparation, avoiding long-term resource occupation |
| Prompt trigger rule | Match against customer specifications | Automatically align with the current bid’s customer requirements to generate documents that meet specifications |
Three Common Misconfigurations
- Symptom: After setting
messagesto retain the first 6 rounds of dialogue, the model still cannot call the batch number mentioned in the previous round. Cause: ThechatIdis not bound to the corresponding bid batch, causing the dialogue context to be reset. - Symptom: Garbled characters appear in generated bid quotation data. Cause: The data encoding format is not specified in the prompt, or non-standard characters in equipment logs are not filtered.
- Symptom: Generated quotation data does not meet preset format requirements. Cause: Fields are not constrained via
variables, and format guidance relies solely on natural language prompts.
How to Confirm the Configuration Is Correct
- Enter a test batch number, initiate the first round of dialogue, then mention the wafer ID again, and check if the model can associate the batch information from the previous round.
- Generate a segment of packaging and testing quotation data, and verify that the output includes the preset
variablesfields. - View the session log to confirm that
chatIdmatches the bid batch ID of the current test. - Simulate a scenario with no interaction for an extended period, and confirm that the session automatically terminates after 300 seconds.
The values provided here are common starting points and should be measured against the reader's own samples.