What Data for This Category Looks Like
Typical entities are process, equipment, quality, and production departments of crystalline silicon cell manufacturers. Daily documents include process cards, diffusion and coating parameters, printing and sintering records, EL detection images, IV grading data, and yield reports. Data sources include silicon wafer incoming materials, MES systems, diffusion furnaces, coating equipment, printing presses, EL and IV testing equipment, and laboratory data. Update frequency follows per production batch, per piece inspection, and daily summary. Document structures include fields such as process parameters, test results, and grading identifiers. Some documents contain raw image data.
Constraints Imposed on Knowledge Base Retrieval and Recall
Multiple heterogeneous data sources require retrieval to adapt to documents in different formats, including structured parameters, unstructured text, and image metadata. Frequently updated production and test data need incremental indexing and real-time recall, to avoid retrieval results lagging behind current production batches due to indexing delays. Professional terminology and detailed fields in documents require high retrieval matching accuracy. Low-relevance redundant results must be filtered, while traceability identifiers such as batch and equipment must be retained. This ensures retrieval results can be mapped to specific production links.
How to Configure
This configuration is based on FastGPT v4.17.0. Specific configuration items and their rationales are listed below:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the association accuracy of solar cell process parameters and test results, filters low-value redundant results |
limit | Top 8 entries | Adapts to the reasonable number of associated documents for a single production batch, avoids context window overflow |
searchMode | fullTextRecall | The query examples emphasize identifiers or specialist terms; test this mode with representative data, not as a product default. |
usingReRank | Enabled | Optimizes the sorting logic of recall results for semantic differences in EL detection images and IV data |
datasetSearchUsingExtensionQuery | Enabled | Adds synonymous expansion of professional terms such as diffusion, passivation, and EL detection, improving recall coverage in professional scenarios |
collectionIds | Configured according to the enterprise's exclusive process library | Limits the retrieval scope to avoid mixing irrelevant business documents |
Three Common Misconfigurations
- Duplicate document chunks may appear in retrieval results, leading to disordered context sequence. The cause is that the custom split chunk order is not retained during index merging, and deduplication verification rules are not configured.
- Access failures may occur when retrieving process documents from external websites. The cause is that access permissions for external data sources are not configured, and retrieval adaptation logic for image documents is not implemented.
- Duplicate training is triggered when importing backed-up knowledge base segments. The cause is that the
collectionIdsconfiguration of the original index is not retained, and directly copying files leads to metadata loss.
How to Verify Successful Configuration
- Submit a query containing diffusion furnace parameters, and check whether the
similarityscores of the recall results fall within the preset range. - Import a batch of duplicate process card documents, and confirm that the retrieval results retain the custom split chunk order and have no abnormal duplicate entries.
- Export the current knowledge base index configuration, and verify that
collectionIdsis limited to the scope of the enterprise's exclusive process library. - Trigger a retrieval test for external data sources, and check whether metadata information for EL detection images can be returned normally.
The values provided are common starting points and should be measured against the reader's own samples.