What data looks like for this category
Daily documents for semiconductor wafer fabrication include process flow cards, batch run records, equipment recipes, SPC charts, exception handling sheets, defect maps, and yield reports. Data sources cover MES systems, equipment FDC modules, defect detection and metrology devices, wafer map files, process recipe libraries, and experimental data. Updates sync with each wafer batch production run, real-time equipment operating status, and daily reports. Document fields include wafer batch numbers, process parameters such as etching power and deposition temperature, defect spatial coordinates, yield proportions, and more. Units include watts, degrees Celsius, microns, and other nanoscale and industrial parameter standards.
What constraints these characteristics impose on knowledge base retrieval and recall
The wafer fabrication process chain is extremely long, with high parameter coupling. Cross-procedure cross-review must cover hundreds of procedures and rework paths. As such, retrieval must support associated matching across cross-module documents. Data updates happen frequently, with strict real-time requirements. The recall link must adapt to real-time synchronized batch records and equipment data, to avoid retrieving outdated process parameters. Subtle drifts in nanoscale parameters may trigger batch defects, so retrieval accuracy requirements are high. Rework windows are limited, so retrieval latency must be controlled within a reasonable range to avoid disrupting production line scheduling. Document types include both structured parameter tables and unstructured maps, so retrieval modes must adapt to matching logic for both text and chart data.
How to set up configurations
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.85–0.92 | Process parameters and defect associations in wafer fabrication are highly correlated. Precise matching of professional expressions and numerical links is required to avoid interference from low-match irrelevant documents during review |
limit | Top 8–12 entries | Covers multi-procedure associated documents, while controlling the context length of a single retrieval to meet the fast decision-making needs of production line review |
searchMode | fullTextRecall | Supports full-text retrieval for cross-procedure documents using the supported fullTextRecall mode. |
usingReRank | Enabled | Reranks initial recall results by process relevance, prioritizing documents directly linked to the current procedure and defect type to improve cross-review efficiency |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands process term queries to cover industry abbreviations and professional customary expressions. For example, expand "etching" to "plasma etching" and "dry etching" to match more comprehensive associated documents |
collectionIds | Specify the collection ID of the target wafer process library | Only retrieve knowledge bases exclusive to this category, avoiding interference from cross-category documents on cross-review accuracy |
Three Common Mistakes
- If cross-procedure retrieval does not return the target process documents, check that the request uses
fullTextRecall, the documents are indexed, and no filters exclude them. - Unable to batch select existing files to add to the temporary retrieval scope. This happens when the multi-select file function is not used. Only single-file sequential addition is supported, leading to extremely low efficiency when filtering 100 files from a set of 500.
- Retrieval requests return timeout or 429 status codes. This stems from failing to configure reasonable retrieval frequency limits, or initiating too many cross-library retrieval requests simultaneously, exceeding system carrying limits. It can also occur if the number of matched documents is lower than the
limitsetting.
How to confirm the configuration is complete
- Verify the request uses
fullTextRecalland returns indexed process documents for known search terms. - Upload batch test files, confirm that multi-select file upload is supported, and multiple documents can be added to the temporary retrieval collection at one time.
- Initiate a retrieval for etching process parameter drift, check that the number of returned results matches the interval set by
limit. - View the retrieval log, confirm that automatic query expansion has taken effect, and matched associated documents containing expanded terms such as "plasma etching" and "dry etching".
- Verify that the similarity of retrieval results falls within the
0.85–0.92interval, meeting preset accuracy requirements.
Note: The configuration values listed above are common starting points. They should be evaluated against specific samples for each use case.