What the data for this category looks like
Business data sources for photovoltaic cell manufacturing include silicon wafer incoming material records, production MES systems, operating parameters of diffusion furnaces, coating equipment, and printing presses, EL and IV test results, and laboratory process documents. Data updates are synchronized per production batch, with real-time updates for single wafer testing, and daily yield summary reports. Daily documents include process cards, diffusion and coating parameter tables, printing and sintering records, EL detection images, IV grading data, and yield analysis reports. Fields cover process temperature, deposition time, electrical performance conversion efficiency, defect type, and other metrics. Units include degrees Celsius, seconds, watts, and others. The configuration on this page references the functional logic of FastGPT v4.17.0.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
The data for this category includes both structured parameters and unstructured detection images and text. Update frequencies cover batch, single wafer, and daily dimensions. This requires the retrieval system to support both precise field matching and full-text content recall. Cross-process process association needs require the retrieval system to connect historical data from texturing, diffusion, passivation, printing, and other links. Single batch data volume is large, so the range of recall results must be limited to avoid context overflow. Data must also be isolated by device, batch, and other dimensions to prevent irrelevant data from interfering with recall accuracy.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the precision requirement for process parameters and defect records, filters low-correlation recall results |
limit | Top 8 entries | Covers typical process association entries for a single batch of photovoltaic cells, controls context length |
searchMode | mixedRecall | Balances structured parameter field matching and content matching for unstructured EL detection images and yield reports |
usingReRank | Enabled | Optimizes the sorting logic of recall results, improves correlation with current queries |
datasetSearchUsingExtensionQuery | Enabled | Supplements context information such as batch number and device model, optimizes recall precision of retrieval keywords |
collectionIds | Configured by business group | Isolates different categories of knowledge base data such as silicon wafer incoming materials, device operation data, and detection data |
Three common mistakes
- An error occurs during training when importing knowledge base data. The interface displays a format error prompt. The cause is mixing old-version exported CSV templates with the current version's field structure, without completing format adaptation.
- The re-ranking function does not work in large model conversations. The API call returns results that are not sorted by relevance, and the function only works normally on the knowledge base test page. The cause is failing to enable the
usingReRankparameter in large model configuration, only configuring the re-ranking option for knowledge base retrieval separately. - The upper limit of knowledge base search citations cannot be adjusted. The number of returned results is fixed and does not meet business requirements. The cause is failing to correctly configure the
limitparameter value, or confusing it with other recall parameters.
How to confirm the configuration is correct
- Upload a single diffusion furnace temperature parameter record, perform retrieval, and check if the recall results include associated EL detection defect data.
- Call the API to retrieve historical process data for a specified device, confirm that the number of returned results matches the configured
limitvalue. - Test adding a batch number as a retrieval keyword, verify that the recall results only include knowledge base data for the corresponding batch.
- After enabling the
usingReRankparameter, compare the result sorting before and after retrieval, confirm that entries with higher correlation are ranked first.
The values provided are common starting points and should be validated against relevant samples.