What the Data for This Category Looks Like
Photovoltaic module encapsulation data originates from cell sorting systems, glass adhesive film and backsheet batch ledgers, workshop MES systems, string welding and lamination equipment logs, EL/IV testers, and insulation withstand voltage test records. Daily generated documents cover layout drawings, material batch records, lamination process parameter tables, EL defect images, IV power test reports, insulation withstand voltage records, and packing lists. Each data entry links to the unique identifier of the corresponding module, and includes batch number, test values, equipment number, and generation timestamp. Updates occur once per module or per batch of materials.
What Constraints Do These Characteristics Impose on the Vector Models and Indexes Workflow?
This category has fine-grained data with high update frequency. New data generates for each module or batch of materials. This requires indexes to support incremental updates to avoid time-consuming full reconstruction. Document types include text parameter tables, visual test images, and structured test records. Vector encoding logic must adapt to different data types to prevent semantic deviation from mixed encoding. Data links to module unique identifiers and batch information. Retrieval must prioritize matching associated fields, so indexes must retain field-level metadata to support precise recall. Process parameter matching accuracy directly impacts production quality, so index recall accuracy must meet business verification requirements.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the conventional text length of lamination process records and IV test reports, avoids including irrelevant cross-batch data in a single chunk |
| Overlap Length | 100–150 characters | Retains contextual association between different chunks of the same module, avoids cutting off key metadata such as batch numbers and equipment numbers |
| Similarity Threshold | 0.75–0.85 | Matches the accuracy requirements of cell current gears, filters low-relevance historical process records |
maxToken | 4096 token | Adapts to the total encoding length of EL image descriptions and test parameters, avoids truncating key test information |
defaultToken | 256 token | Used for vector encoding of short single fields, such as batch numbers and equipment numbers in packing lists |
| Reranking | Enabled, top 3 results | Prioritizes recalling test records most relevant to process questions, adapts to high-frequency retrieval scenarios such as EL crack detection and power matching |
This configuration is based on FastGPT v4.17.0.
Three Common Misconfigurations
- Phenomenon: Index model integration fails during Docker intranet deployment, with logs showing connection timeout. Cause: The index model service port is not mapped to the host intranet segment, so the FastGPT instance cannot access the model service.
- Phenomenon: The exported knowledge base dataset.csv only contains the index field, with no content field. Cause: The text extraction switch is not enabled during document import, so only document metadata is stored, and no main text content is extracted.
- Phenomenon: When using the bge-large model, long texts are imported but do not generate vectors as expected after segmentation. Cause: The chunk length parameter is not adjusted, and the default segmentation logic does not match the structure of long process records, resulting in failed segmentation.
How to Verify Correct Configuration
- Perform an incremental import of a single lamination process record, check the index update log, confirm that only data for the corresponding batch is added, and no full reconstruction prompt appears.
- Initiate a retrieval request for cell current gear matching, verify the matching degree between the associated fields of the returned results and the query conditions, adjust the similarity threshold to the range that meets business requirements.
- Export the knowledge base dataset.csv, check whether both index and content fields are included, confirm that the text extraction switch is correctly configured.
- Check GPU resource usage, confirm that multiple graphics cards are scheduled for index construction and vector encoding, with no idle resources.
The values provided here are common starting points and should be evaluated against the reader's own samples.