What data looks like for this category
Data sources for this category include wafer test maps, MES systems, operation logs of bonding and molding equipment, output results from ATE test machines, X-ray/SAM inspection reports, and customer specification documents. Data is updated per wafer batch production, per individual chip test, and generated daily. Daily documents include packaging process cards, bonding and molding parameter sheets, incoming wafer maps, test program version files, yield reports, and failure analysis reports. Data fields include batch numbers, unique chip IDs, test parameter values, inspection results, equipment numbers, and some parameters with physical units such as soldering temperature, molding pressure, and test voltage.
Constraints on multi-turn dialogue and prompt engineering
High-frequency, fine-grained data updates require multi-turn dialogues to synchronize the latest test parameters and inspection results for the current batch in real time, and avoid using expired process cards or test program versions. Diverse document types with significant structural differences require prompt engineering to adapt to the field formats of different documents, such as distinguishing coordinate fields from incoming wafer maps and statistical fields from yield reports. Single data entries are linked to multiple equipment and process nodes, so multi-turn dialogues must associate context using batch numbers to avoid mixing data across batches. Additionally, data sources include sensitive logs from production equipment, so interface permissions called by dialogues must match the access scope of equipment data to prevent unauthorized access.
How to configure settings
This configuration applies to FastGPT v4.17.0.
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
chatId | Generate a unique identifier per production batch | Ensure context association for inspection dialogues in the same batch, avoid mixing data across batches |
messages | Carry the first 3 rounds of inspection records and basic parameters of the current batch in each dialogue | Cover key process nodes of the batch, while controlling context length to avoid overflow |
variables | Preset two variables: batch number and test program version | Quickly locate documents and equipment data for the corresponding batch, improve retrieval accuracy |
| Single search limit | Top 5 entries | This category has a large number of data entries per batch, limiting search entries avoids redundant information interfering with dialogue results |
| Maximum context length | 800–1200 characters | Adapt to the document field length of semiconductor packaging and testing, balance context completeness and response speed |
| Question optimization | Calibrate based on actual testing | Inspection process communication scripts vary significantly across factories, so adjust optimization logic based on actual scenarios |
Three common configuration errors
- Symptom: Returns empty fields or insufficient result counts when attempting to pull all dialogue logs. Cause: Failed to correctly configure the association rules for
chatId, or did not enable the persistent storage switch for dialogue data. - Symptom: Wafer test maps or incoming wafer maps cannot be correctly interpreted after upload. Cause: Did not specify target fields for image parsing in the prompt, or did not enable the corresponding configuration for image content extraction.
- Symptom: Dialogue context does not include historical records, and search results do not match the current batch’s inspection requirements. Cause: Did not pass the
messagesparameter in the order of dialogue turns, or set the context length too small, resulting in historical records being truncated.
How to verify successful configuration
- Initiate an inspection dialogue for the current production batch, check whether the returned results include packaging process card and test program version information for the corresponding batch.
- Attempt to upload an incoming wafer map, verify whether the dialogue can extract chip coordinate and batch identification fields from the image.
- Initiate 3 consecutive dialogues associated with the same batch, check whether subsequent replies retain key parameters from previous dialogues.
- View the dialogue log module, confirm that all dialogue records under the corresponding
chatIdcan be exported.
The values provided are common starting points and should be measured against samples relevant to the specific deployment scenario.