What This Category of Data Looks Like
Data sources for this category include MES systems, equipment FDC collection modules, defect detection and metrology equipment, process recipe libraries, and more. Daily documents cover process flow cards, batch run records, equipment recipes, SPC charts, exception handling sheets, defect maps, and yield reports. These documents include fields such as nanoscale process parameters, wafer batch numbers, defect spatial coordinates, and yield ratios. Data update rhythm follows wafer batch progression: key equipment parameters are collected in real time, and regular reports are updated daily.
Constraints on Citation and Traceability Processes
The characteristics of ultra-long process chains coupled with nanoscale parameters require that citation traceability be precisely bound to the corresponding wafer batch and process node, to avoid cross-batch parameter confusion. Real-time collected equipment data coexists with daily updated yield reports: the priority of data sources with different update frequencies must be distinguished to ensure traceability points to the latest valid version. Cleanroom rework windows are limited, so traceability response speed must match on-site decision rhythm. Citations of exception handling sheets must be linked to corresponding defect maps to avoid attribution bias.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
quoteMaxToken | 1500–2000 token | Matches the average token count of a single wafer fabrication document, avoids context overflow |
responseData | Enabled | Citation source information of referenced documents must be displayed on the official chat page to meet business traceability requirements |
detail | Full fields | Exclusive fields such as document batch number and affiliated system must be displayed for on-site personnel to verify |
| Search Filter | Filter by wafer batch | Aligns with the batch-based business logic of wafer fabrication management, narrows traceability scope |
| Single Search Limit | Top 3–5 entries | Core process attribution only requires a small number of key documents, avoids interference from irrelevant information |
| Maximum Context | 8000–12000 token | Matches the total token count of documents for a single wafer batch, avoids truncation of critical parameters |
Three Common Configuration Mistakes
- Phenomenon: Citation source information of referenced documents is not displayed on the official chat page, and only visible in the debug page. Cause: The
responseDataconfiguration item is not enabled, and traceability data is only retained in debug mode. - Phenomenon: The number of results returned by knowledge base search matches the set value, but the actual token count exceeds the threshold. Cause:
quoteMaxTokenis not set reasonably, only the number of returned entries is limited, not the token occupancy of a single document. - Phenomenon: Program results executed in the workflow cannot be used as background knowledge input to subsequent modules. Cause: The output fields of the execution results are not bound to the context input node of the knowledge base, and data flow configuration is not completed.
How to Verify Successful Configuration
This content is written based on the FastGPT v4.17.0 version configuration logic. The following actions can be used to verify configuration effects:
- Enter the officially released chat page, initiate a query related to wafer batch parameters, check whether the source path and exclusive fields of the referenced documents are displayed below the reply.
- Adjust the
quoteMaxTokenconfiguration item to different intervals, initiate a query containing multiple process documents, verify whether the total token count of the returned referenced documents matches the set value. - Configure the search filter rule by wafer batch, enter a specific batch number, check whether the returned documents are only associated with the run records and reports of that batch.
- Enter the workflow configuration page, bind the output results of program execution to the context input node of the knowledge base search, run the workflow to verify that data flow is normal.
The values provided are common starting points and should be measured against the reader's own samples.