What the Data for This Category Looks Like
Data from the photovoltaic module encapsulation process comes from production departments such as string welding, lamination, and testing. It includes layout diagrams, material batch records, lamination process parameter tables, EL crack images, IV test reports, insulation voltage resistance records, and packing lists. Data sources include the battery piece grading system, glass adhesive film backplane batch ledger, factory MES system, string welding and lamination device logs, and EL/IV tester export files. The data update frequency is one set of test records per module, and one batch ledger per batch of raw materials. Each document includes fields such as batch number, current value, test time, and device number, with units including ampere, volt, pascal, and other professional parameters.
What Constraints These Characteristics Impose on the Vector Model and Indexing Link
The business data of photovoltaic module encapsulation has the characteristics of high-frequency updates, mixed multi-formats, and close business association. High-frequency updates require indexes to support incremental synchronization, avoiding resource consumption caused by full reconstruction. Multi-format documents (such as transcribed text from EL images, tabular process parameters) require chunking rules adapted to the granularity of different content, avoiding truncation of critical information across business blocks. The strong association between batches and modules requires indexes to retain contextual association, ensuring that full-link data for the same batch can be matched during retrieval. At the same time, the units and fields of professional parameters require the vector model to accurately recognize the semantic association of professional terms, avoiding misjudgment.
Configuration Settings
The following configurations are formulated based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the length of a single page of a single module test report, avoids truncating process parameters or batch information |
| Overlap Length | 100–200 characters | Retains contextual association between batch records and corresponding modules, avoids losing association after chunking |
defaultToken | 8192 | Adapts to the upper limit of text length after EL image transcription, ensures long text chunks can be vectorized normally |
maxToken | 16384 | Covers the total length of multiple batch documents for batch indexing, avoids exceeding model limits during single indexing |
| Similarity Threshold | 0.75–0.85 | Distinguishes qualified and unqualified EL crack judgment results, ensures retrieval accuracy meets quality audit requirements |
| Single Index Limit | 50 items per batch | Avoids excessive server load during batch indexing, adapts to resource configurations in production environments |
Three Common Errors
- When uploading EL images or IV test reports larger than 10MB, some chunks experience vectorization anomalies and cannot recover after repeated retries. The cause is that a unified chunking rule is not configured for multi-format documents, and some long text chunks exceed the upper limit set by
defaultToken, leading to vectorization failure. - Batch indexing a large number of batch records causes server crash, and the knowledge base cannot automatically become ready and gets stuck after restart. The cause is that breakpoint resume configuration is not enabled, and unfinished indexing tasks are not resumed after restart, requiring manual re-triggering of indexing.
- After updating the FastGPT version, the knowledge base with the original index cannot support vector retrieval, and no results are returned for searches. The cause is that the embedding dimension of the vector model changes after version upgrade, and the vector format of the original index is incompatible with the new configuration.
How to Confirm the Configuration Is Properly Set
- Upload a typical lamination process record document, verify that the number of chunked text blocks matches the business logic, and no critical parameters such as lamination temperature and time are truncated.
- Run a batch indexing test, index 50 batch records in a single batch, confirm that server load remains within a reasonable range, with no timeout or error logs generated.
- Search for preset battery current gear keywords, confirm that the similarity of returned results falls within the set threshold range, with no irrelevant content included.
- Restart the FastGPT service, check that the index status of the knowledge base automatically recovers to ready, without requiring manual triggering of indexing tasks.
The values provided here are common starting points and should be measured against the reader's own samples.