What the data for this category looks like
The service subjects for this category are the process integration, module, equipment, yield, and manufacturing departments of wafer fabs. Daily documents include process flow cards, batch run records, equipment recipes, SPC charts, exception handling sheets, defect maps, yield reports, and more. Data is primarily sourced from MES systems, equipment FDC, defect detection and metrology devices, wafer maps, process recipe libraries, and experimental data, with update frequencies covering per wafer batch, real-time, and daily. Documents contain numeric process parameters, graphical inspection results, and structured batch information, with most fields carrying unique identifiers such as nanoscale parameters and batch numbers.
What constraints these characteristics impose on knowledge base retrieval and recall
Multiple heterogeneous data sources require retrieval to support associated recall of cross-system data, avoiding siloed process information. The high-frequency update rhythm requires the indexing mechanism to quickly synchronize the latest documents such as equipment recipes and exception handling sheets, preventing the use of outdated process parameters during bidding. The long process chain and nanoscale parameter characteristics require retrieval to accurately match process details, avoiding generalized results that interfere with the rigor of bid documents. The coexistence of graphical documents and structured data requires retrieval tools to support content parsing across different formats, while ensuring the accuracy of parameter comparison. In addition, the bidding scenario requires rapid access to compliant historical data, which places clear requirements on retrieval response speed.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the precision requirements for wafer fabrication process parameters, filtering out low-correlation non-process documents |
limit | Top 8–12 entries | Covers associated data across multiple processes, while controlling the context length of retrieval results to adapt to the typesetting requirements of bid documents |
searchMode | mixedRecall | Matches both process keywords and numeric parameters, adapting to retrieval of mixed-format documents such as SPC charts and equipment recipes |
usingReRank | Enabled | Reranking optimizes the ranking of process-related documents, preventing superficially similar but process-irrelevant results from appearing in top positions |
datasetSearchUsingExtensionQuery | Enabled | Expands the search scope to recall associated documents relevant to bidding topics, such as yield reports and exception handling sheets |
collectionIds | Bind corresponding collection IDs by modules such as process integration, equipment, and yield | Distinguishes exclusive documents for different departments, preventing retrieval of irrelevant cross-category data |
Three Common Errors
- Garbled characters appear after importing Chinese process documents. The cause is that the original document encoding is not set to UTF-8. Locally exported documents for wafer fabrication often use non-standard encoding formats.
- Newly added equipment recipe documents cannot be retrieved. The cause is that incremental index updates for the knowledge base were not triggered. The update frequency of documents for this category is per wafer batch or real-time; failure to update in a timely manner will prevent access to the latest data during bidding.
- The number of retrieval results is far lower than the set
limitvalue. The cause is that thesimilaritythreshold is set too high, filtering out low-similarity but relevant documents that meet process matching requirements.
How to Confirm the Configuration Is Properly Set
- Upload a latest batch run record document, wait for FastGPT v4.17.0 to complete indexing, search for the unique batch number, and confirm that the corresponding document can be retrieved.
- Import a Chinese-encoded process card document, check that the page displays no garbled characters, and confirm that the knowledge base encoding setting has been adjusted to UTF-8.
- After enabling
usingReRank, compare the ranking of retrieval results before and after enabling, confirm that the reranking function takes effect, and that the ranking of process-related documents is more reasonable. - Check the disk usage of the locally deployed knowledge base, confirm that the statistical scope includes original files, split file chunks, and embedding vectors, which meets disk management requirements.
The values provided here are common starting points and should be evaluated against the reader's own samples.