What the data for this category looks like
The inspection and process record data for photovoltaic module encapsulation comes from multiple business carriers across the string welding, lamination, framing, and testing departments. Daily documents include layout drawings, material batch records, lamination process records, EL crack images, IV test reports, insulation voltage withstand records, and packing lists. Data sources cover battery cell grading results, glass adhesive film backplane batch ledgers, MES systems, string welding and lamination device logs, and EL/IV tester outputs. Data is updated to generate exclusive test records for each module, and batch ledgers for each batch of raw materials. Individual records include fields such as material batch numbers, process parameters, test values, and image files.
Constraints on Vector Models and Indexes Workflow
This category has high data update frequency and scattered per-item volume. Vector indexes must support incremental updates to adapt to batch component audit scenarios. Data includes both structured text (process records, test reports) and unstructured images (EL images). Vector models must support embedding for both data types. Matching battery cell current grades is a core time-consuming step. Vector recall accuracy must meet the tolerance matching requirements of process parameters to avoid misjudgment or missed judgment. Format differences across multiple data sources require the index layer to unify data embedding and storage formats, reducing matching costs for subsequent retrieval.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the single-segment content length of EL image text annotations and lamination process records, avoids splitting critical information within a chunk, optimized based on FastGPT v4.17.0 chunking logic |
| Overlap Length | 100–150 characters | Covers connected content for process steps such as string welding and lamination, prevents loss of process-related information across chunks |
maxToken | 8192 token | Adapts to the structured text length of IV test reports, meets the single-sample token limit requirements of mainstream multimodal embedding models |
| Similarity Threshold | 0.75–0.85 | Meets the accuracy requirements for matching battery cell current grades, can be adjusted and calibrated based on the actual tolerance range of components |
| Single Search Limit | Top 10 results | Balances recall efficiency and result coverage for batch component audits, avoids excessive redundant results that reduce audit efficiency |
| Reranking | Enabled | Optimizes the sorting logic for high-similarity results, prioritizes key reference records for lamination bubble and crack judgment |
Three Common Misconfigurations
- Phenomenon: After uploading EL crack images, the vector embedding task returns empty results or format errors. Cause: A multimodal vector model is not configured, and only a plain text embedding model is used, which cannot meet the embedding requirements for image data.
- Phenomenon: The file upload API call returns success, but subsequent retrieval has no matching results. Cause: FastGPT v4.17.0 incremental index asynchronous completion is not waited for, and a retrieval request is sent directly. The index has not been synchronized yet.
- Phenomenon: Recall results deviate significantly when matching battery cell current grades, leading to incorrect matches. Cause: The similarity threshold setting does not consider the current tolerance range of photovoltaic modules. A threshold that is too high will filter valid matching results, while a threshold that is too low will introduce a large number of irrelevant records.
How to Verify Correct Configuration
- Upload a single lamination process record, check the system embedding task log, confirm there are no abnormal errors, and the generated vector data can be stored normally.
- Initiate a retrieval request for battery cell current grade matching, verify that the returned result relevance conforms to the preset matching logic, and adjust the similarity threshold for verification.
- Call the file status query interface provided by FastGPT v4.17.0, confirm that both the original uploaded file data and the index are completed before initiating formal retrieval.
- Count the token consumption of batch test data, confirm that the
defaultTokenandmaxTokenconfigurations do not exceed the limit requirements of the selected embedding model.
The values provided are common starting points and should be measured against the reader's own samples.