What the data for this category looks like
Data typically comes from silicon wafer incoming material records, production MES systems, operation data from diffusion furnaces, coating equipment, or printing presses, EL and IV test results, and laboratory analysis data. Updates occur in real time for each production batch and single-chip test result. Production and quality data are aggregated daily. Daily documents include process cards, diffusion and coating parameter tables, printing and sintering records, EL detection images, IV grading data, and yield reports. These documents contain fields such as batch number, process parameter values, test results, and grading levels. Units include voltage, current, and conversion efficiency percentage, among others.
What constraints these characteristics impose on knowledge base retrieval and recall
Multi-source heterogeneous data sources require the retrieval module to support associated recall across different systems. High-frequency updated production and test data needs adaptation to real-time or near-real-time retrieval refresh mechanisms. Diverse document types — structured parameter tables, visual detection images, text reports — require compatibility with content parsing and vector modeling for different formats. Single-batch production has strong correlation. Retrieval must prioritize recalling process and test data associated with the same batch. This ensures accuracy of exception tracing. High precision requirements for process parameters mean retrieval matching precision must adapt to narrow numerical ranges.
How to configure the settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets the high precision requirements for photovoltaic process parameters, and avoids recalling irrelevant batch data |
limit | Top 8 entries | Balances retrieval efficiency and recall coverage, ensures sufficient context for exception report drafting |
searchMode | mixedRecall | Combines semantic and keyword matching, supports mixed retrieval of structured parameter tables and unstructured detection reports |
usingReRank | Enabled | Re-ranks initial recall results, prioritizes core process parameters corresponding to exception phenomena |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonyms for photovoltaic professional terminology, adapts to industry-specific expressions |
collectionIds | Exclusive knowledge base collections bound to production, equipment, and quality departments | Isolates documents from different departments, avoids interference from cross-department retrieval |
Three common errors
- Symptom: A 404 status code is returned when calling knowledge base retrieval. Cause: The specified knowledge base collection is not correctly bound, or the initialization script is not executed under the corresponding FastGPT version.
- Symptom: Retrieval results are empty or contain no matching process parameters. Cause:
datasetSearchUsingExtensionQueryis not enabled, so synonym expansion for photovoltaic professional terminology cannot be performed. - Symptom: The number of retrieval results exceeds expectations. Cause: The
limitparameter is not set correctly, or a recall upper limit that exceeds server resource capacity is configured.
How to verify successful configuration
- Check the system logs for knowledge base retrieval, confirm that returned results include process and test documents from the target department.
- Simulate retrieval for an exception scenario, verify whether results prioritize recalling associated production data from the same batch.
- Adjust the
similarityparameter, confirm that matching precision adapts to process parameter retrieval requirements through multiple test sets. - Confirm that the currently deployed FastGPT version is v4.17.0, which matches the compatibility requirements for the knowledge base configuration.
The values provided here are common starting points, and should be evaluated against the reader's own samples.