What the Data for This Category Looks Like
Business data for this category comes from wafer fab MES systems, equipment FDC modules, defect detection and metrology equipment, wafer image libraries, process recipe libraries, and experimental data platforms. Daily documents include process flow cards, batch operation records, equipment recipes, SPC charts, exception handling sheets, defect maps, and yield reports. Data update frequencies cover per wafer batch, real-time, and daily. Fields include wafer ID, process step number, parameter thresholds, defect spatial coordinates, and more. Units include nanometers, pascals, batch numbers, and more.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
The long process chain creates a large volume of multi-process documents. Retrieval must limit the number of recalled results to avoid exceeding the context window. Heterogeneous data from multiple sources (text recipes, numerical SPC charts, wafer maps, etc.) requires retrieval modes that balance semantic and keyword matching. High-frequency data updates require indexes to support incremental synchronization, ensuring data timeliness. The nanoscale parameter coupling business characteristic requires recalled results to have high matching accuracy, avoiding irrelevant process data. The cross-module yield attribution business requirement requires retrieval to associate data sources across multiple business sets.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75-0.85 | Match nanoscale process parameters and exception descriptions to avoid mistakenly recalling irrelevant process data |
limit | Top 8-12 entries | A single wafer batch covers a wide range of process steps. Too many results exceed the context window, while too few fail to cover associated exceptions |
searchMode | mixedRecall | Balance semantic matching for text recipes and keyword matching for SPC charts and defect maps |
usingReRank | Enabled | In multi-source heterogeneous data scenarios, reranking improves the ranking priority of highly relevant results |
datasetSearchUsingExtensionQuery | Enabled | Equipment inspection questions often imply parameter drift and recipe changes. Extended queries can associate implicit business-related data |
collectionIds | Configured by business module grouping | Accurately associate corresponding data sources such as equipment FDC, yield reports, and process recipes, and support cross-module attribution |
Three Common Configuration Errors
- Phenomenon: The exported knowledge base dataset CSV file only contains the index field and no content field. Cause: The option to include original text content during export is not checked, or correct data source table fields are not associated during custom index configuration.
- Phenomenon: The page keeps refreshing when importing documents to the knowledge base, and the target page cannot be selected. Cause: Imported documents contain unparsed nested format elements, or the total number of documents imported in a single batch exceeds the system temporary cache threshold, causing page rendering blocking.
- Phenomenon: After uploading a knowledge base file in an offline environment, the status keeps spinning, and normalcy resumes after connecting to the internet. Cause: The system accesses external resources to complete dependency verification during initialization. The verification process cannot complete in offline status, causing the upload process to block.
How to Verify Correct Configuration
- Run a query targeting equipment parameter drift, and check whether the similarity scores of returned results fall within the preset 0.75-0.85 range.
- View the index construction log to confirm that multi-source data such as MES records, equipment FDC data, and defect maps have been correctly associated with corresponding business sets.
- Export the test dataset to confirm that the content field fully contains original text content with no missing fields.
- Disconnect the network and upload a small test document to confirm that the upload process completes normally with no persistent spinning blocking.
The values provided are common starting points and should be evaluated against the reader’s own samples. This page’s configurations are based on FastGPT v4.17.0.