What the data for this category looks like
Business data for semiconductor wafer fabrication originates primarily from MES systems, equipment FDC modules, defect detection and metrology tools, wafer map files, process recipe libraries, and experimental datasets. Data updates align with three cycles: each wafer batch production run, real-time equipment parameter synchronization, and daily report aggregation.
Daily documents include process flow cards, batch operation records, equipment recipe parameters, SPC control charts, exception handling sheets, defect maps, and yield reports. Most fields contain nanoscale process parameters, batch unique identifiers, run durations, and defect spatial coordinates. Data structures cover two categories: structured parameter tables, and unstructured images and text.
What constraints these characteristics impose on the knowledge base retrieval and recall link
This category has high-frequency data updates, plus both structured parameters and unstructured images and text. This requires retrieval systems to support incremental synchronization and multi-modal content recall.
The combination of ultra-long process chains and nanoscale parameters demands precise matching of batch identifiers and process parameters during retrieval, to avoid interference from irrelevant data. Business requirements for limited rework windows mean retrieval response times must stay within a reasonable range.
The need for cross-module yield attribution requires retrieval to filter datasets by process module, narrowing the recall scope to improve accuracy. The need to interpret defect spatial distribution requires retrieval to associate semantic information across different document types. Matching only a single field cannot cover this requirement.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets precision requirements for matching nanoscale process parameters and defect descriptions, avoids false or missed recalls |
limit | Top 8–12 entries | Covers associated data across hundreds of process steps, while controlling the context length of a single retrieval |
searchMode | mixedRecall | Balances structured parameter matching and semantic recall for unstructured images and text |
usingReRank | Enabled | Performs secondary ranking on initial recall results, optimizes semantic matching for defect spatial distribution and process paths |
datasetSearchUsingExtensionQuery | Enabled | Expands synonymous terms for process terminology, covers different naming conventions for the same process across departments |
collectionIds | Collection IDs divided by process module | Splits datasets along process chains, narrows cross-module retrieval scope and improves yield attribution accuracy |
Three common errors
- Symptom: Retrieval results include non-target batch wafer operation records, and irrelevant data cannot be filtered by batch. Cause: Metadata filtering rules for the dataset are not configured, and
collectionIdsare not divided by process module. This leads to recall scope covering the full dataset. - Symptom: No response is returned for a long time after initiating retrieval, and a timeout error is displayed. Cause: A reasonable
limitparameter is not set, andusingReRankis not enabled. This leads to too many documents being recalled in a single run, exceeding system processing limits. - Symptom: Retrieval results for defect maps do not link to corresponding exception handling sheets, and semantic matching performance is poor. Cause:
datasetSearchUsingExtensionQueryis not enabled, so synonymous process terminology cannot be expanded. This causes failed associated recall across different documents.
How to confirm the configuration is properly set
- Log in to the FastGPT v4.17.0 admin backend, navigate to the configuration page of the target knowledge base, confirm that
collectionIdsare divided by process module, and that metadata filtering rules enable the batch number field. - Enter the process parameters of a specific wafer batch as a retrieval query, verify that the update times of returned results match the per-batch, real-time, or daily update cycles, and that only documents from the target process module are included.
- Check the retrieval configuration items, confirm that
similarity,limit,usingReRank,datasetSearchUsingExtensionQueryand other parameters are set according to actual measurement calibration. - Enter keywords for defect maps, verify that returned results include corresponding exception handling sheets and yield reports, to validate the effect of multi-document associated recall.
The values provided are common starting points and should be measured against the reader's own samples.