What data for this category looks like
Data for photovoltaic module encapsulation comes from the battery sheet grading system, glass adhesive film and backplane batch ledgers, workshop MES system, series welding and lamination equipment operation logs, and EL/IV tester export files. Daily retained documents include equipment layout drawings, material batch records, lamination process parameter tables, EL crack images, IV test reports, insulation withstand voltage test records, and finished product packing lists. Data is updated as follows: one test record is created per component, a batch ledger is generated synchronously for each raw material batch, and each single data entry links device operating parameters and product quality inspection results.
Constraints imposed by these characteristics on vector models and indexing
Data includes structured ledgers, image files, and time-series logs. Mixed-type data requires indexes to support multimodal association and cross-document context matching. High-frequency updated single-component-level data increases incremental synchronization pressure on indexes. Indexing strategies adapted for small-batch high-frequency writes are required. EL images and IV test reports have closely linked fields. Cross-document context association capability must be retained to avoid separating the correspondence between parameters and images during retrieval. The time-series nature of device logs requires indexes to support filtering by time range. This matches historical data query needs in equipment inspection scenarios.
How to set configurations
The following configurations are developed for the vector model and indexing features in FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the single-section parameter length of lamination process records and IV test reports, avoiding splitting text blocks that span critical parameters |
| Chunk Overlap Length | 100–150 characters | Retains context continuity between adjacent chunks, ensuring continuous text related to EL image interpretation is not split |
| Similarity Threshold | 0.72–0.85 | Matches precise retrieval needs for equipment inspection scenarios, filtering low-relevance batch records |
| Maximum Search Results Per Query | Top 8 entries | Adapts to lightweight retrieval for maintenance Q&A, avoiding excessive redundant device log results |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
defaultToken | Fill in according to platform configuration | Ensures valid identity verification for vector model calls, compatible with both local and cloud deployment scenarios |
Three common configuration mistakes
- The interface displays "Training order creation failed" with a 400 status code. The cause is failure to correctly link photovoltaic module encapsulation-specific material batch records and device log data sources, resulting in missing associated fields during index construction.
- Retrieval results only return text data and do not link to corresponding EL images. The cause is failure to enable multimodal index configuration and include EL image files in the vector model’s vectorization scope.
- A prompt stating "Vector model unavailable" appears during local deployment, even though oneAPI is connected. The cause is failure to configure
defaultTokenas a valid token for the corresponding Embedding model, or failure to specify the retrieval scope for the photovoltaic module scenario in collection settings.
How to confirm the configuration is correct
- Upload a combined document of lamination process records and EL images, run a retrieval test, and check if returned results include linked batch records and device logs.
- View the index construction log to confirm single data entry chunk lengths match the preset range, with no abnormally split critical parameter text.
- Verify the vector model’s token configuration, ensuring
defaultTokenmatches the token configured in the calling interface. - Run a retrieval using fault keywords, and check if returned results are sorted to prioritize high-frequency fault scenarios in equipment inspection.
The values provided are common starting points and should be measured against the reader's own samples.