What This Category’s Data Looks Like
Data for the photovoltaic module encapsulation process comes from battery sheet grading ledgers, glass adhesive film and backplane batch records, MES systems, string welding and lamination equipment operation logs, EL tester output images, IV test reports, insulation voltage withstand records, and other sources. Daily documents include production layout diagrams, material batch tables, lamination process parameter tables, EL defect images, packing lists, and more. Data is updated in real time for each module and each material batch. Fields include battery sheet current gear, lamination temperature and pressure parameters, EL crack locations, power test values, and additional metrics. Document structure varies slightly based on the output format of production equipment.
Constraints Imposed on Knowledge Base Retrieval and Recall
Data for this category is updated frequently per individual module and single batch. Single record sizes vary significantly. The dataset includes structured batch fields alongside unstructured images and test reports. The retrieval link must support both exact matching of structured fields and full-text recall of unstructured documents. This prevents traceability deviations caused by delayed data updates. Fields include quality-critical content such as battery sheet current gear and lamination process parameters. Recalled content must be accurately bound to the batch and process parameters of the module being traced. This stops irrelevant data from interfering with analysis results. Cross-department documents must be categorized by production processes such as string welding and lamination. This ensures the retrieval scope focuses on the production process linked to the current issue.
Configuration Settings
These configurations are adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Matches the matching accuracy of structured batch fields and unstructured test reports for photovoltaic module encapsulation, balances recall accuracy and coverage, adjust based on actual testing |
limit | Top 8–12 entries | Matches multi-dimensional data (batch, process, test) for a single module, avoids excessive redundant data exceeding the context window, while covering enough associated records |
searchMode | mixedRecall | Supports both structured field exact matching and unstructured full-text retrieval, adapts to the mixed-format document data of this category |
usingReRank | Enabled | Reorders recall results by relevance to the query, optimizes the display order of key content such as EL cracks and power test results, improves analysis efficiency |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands related process terms based on the user's traceability query, associates lamination process records with material batch data |
collectionIds | Knowledge base collection IDs bound to production processes | Limits the retrieval scope to knowledge bases of departments corresponding to string welding, lamination, testing, etc., avoids interference from irrelevant cross-process data |
Three Common Configuration Mistakes
- Phenomenon: A knowledge base API call returns an unauthorized error, even when a valid API key is configured. Cause: The current application is not bound to the corresponding photovoltaic module encapsulation knowledge base collection, or the collection’s access permissions are not granted to the current API key.
- Phenomenon: After adjusting the
limitparameter to 2000, the large language model does not reference any recalled content, while the retrieval details show data was successfully recalled. Cause:usingReRankis not enabled, leading to disordered recall result sorting, or the context window is exceeded, causing valid content to be automatically truncated and unreadable by the large language model. - Phenomenon: A knowledge base built from local documents has poor answer quality, and cannot accurately match battery sheet current gears and power test results. Cause: A reasonable range for the
similarityparameter is not set, leading to a large number of irrelevant unstructured documents in recall results, or knowledge base collections are not divided by production process, leading to an overly broad retrieval scope.
How to Confirm Proper Configuration
- Submit a retrieval query for a specific component batch, and verify that the recall results include the batch’s material ledger, lamination process records, and EL test reports.
- Check the recall result list returned by the API, confirm that the number of recalled entries controlled by the
limitparameter matches the preset range, and that the relevance of results matched by thesimilarityparameter meets business expectations. - After enabling
usingReRank, verify that the reordered results place the most relevant content to the current query (such as process parameters corresponding to lamination delamination) at the top. - Check the application's permission configuration, confirm that the bound knowledge base collection ID is correct, and that the API key has access permissions for the corresponding collection.
The values provided are common starting points and should be measured against the reader's own samples.