What the Data for This Category Looks Like
Data sources for photovoltaic module encapsulation scenarios include the wafer grading system, glass adhesive film and backplane batch ledger, workshop MES system, string welding and lamination equipment, EL and IV testers. Daily generated documents include layout drawings, material batch records, lamination process records, EL images, IV test reports, insulation withstand voltage records, and finished product packing lists. Data updates occur each time a module is produced or a batch of raw materials is put into storage. Each single data entry is linked to the raw material parameters of the corresponding batch and process indicators from production links including string welding and lamination.
Constraints on Vector Models and Indexes From These Data Characteristics
The data characteristics of this category impose multiple constraints on the vector model and index workflow. High-frequency, fine-grained updates require indexes to support incremental synchronization, avoiding performance losses from full reconstruction. Diverse document types cover structured ledgers and unstructured images and drawings, requiring vector models to support embedding of both text and image features. Each single data entry is linked to precise business indicators such as wafer current grades and lamination process parameters. This requires the similarity threshold for index recall to match the accuracy requirements of process matching, preventing irrelevant data from interfering with tender quotation accuracy. Multi-dimensional production and raw material tags require indexes to support multi-condition associated retrieval by batch and production link.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the single-paragraph text length of process records and test reports for photovoltaic module encapsulation, avoiding information fragmentation within chunks |
| Overlap Length | 100–150 characters | Covers cross-chunk process parameter associations, preventing key information such as wafer current grades from being split across chunk boundaries |
maxToken | 8192 token | Adapts to the total length of IV test reports and EL image annotation text, ensuring complete embedding of key parameters |
| Similarity Threshold | 0.75–0.85 | Matches the matching accuracy requirements of wafer current grades, filtering low-correlation historical project data |
| Single Search Limit | Top 10 entries | Controls the volume of retrieval results, adapting to the need for fast matching of historical cases for tender documents |
defaultToken | Calibrated via actual testing | Adapts to embedding requirements for different batches of raw materials, avoiding feature deviations caused by fixed values |
Three Common Misconfigurations
- Phenomenon: The interface prompts "This token does not have permission to use the model: text-embedding-3-large" or reports an error with no available vector channel. Cause: The configured vector channel permissions are not synchronized to the FastGPT token pool, or the key configured for the channel is not bound to the corresponding model permissions.
- Phenomenon: Knowledge base search takes more than 30 seconds, and single-batch data retrieval has no response. Cause: Incremental indexing is not enabled. Full indexing requires traversing all photovoltaic module production data for each retrieval, or the chunk length is too large, leading to excessive embedding calculation time per chunk.
- Phenomenon: Non-corresponding batch wafer parameter data is mixed in retrieval results. Cause: The similarity threshold is set too low, failing to filter historical data that does not match the current tender project parameters, or index tags are not added according to raw material batches.
How to Verify Proper Configuration
This page is based on FastGPT v4.17.0.
- View the vector model embedding logs to confirm that the embedding time for single photovoltaic module process records meets expectations, with no timeout errors.
- Manually enter the current grade parameter of a batch of wafers, confirm that the retrieval results prioritize returning historical project data of the same grade, to verify the similarity threshold matching logic.
- Add a test packing list data entry, confirm that the index automatically completes incremental synchronization, with no full reconstruction prompt.
- Check the FastGPT token configuration page, confirm that the bound vector channels include the currently used model, with no access denied prompts.
The configuration values listed above are common starting points. Adjust them based on your own specific samples and requirements.