What Data Looks Like for This Category
Routine documents for this category include layout diagrams, material batch records, lamination process records, EL images, IV test reports, insulation voltage withstand records, and packing lists. Data sources cover cell grading, glass adhesive film backplane batch data, MES systems, data collected from string welding and lamination equipment, and output data from EL/IV testers. Records are generated for each module and each batch of materials. Documents include fields such as batch number, current gear, process parameters, and test values. Some graphic and text documents must be linked to text records for the corresponding batch.
Constraints Imposed on Vector Models and Indexes
High-frequency updates for each module and batch of materials require indexes to support low-latency writes and queries, ensuring real-time cross-review. Mixed storage of multiple document types requires indexes to adapt to vector mapping logic for different formats, including text, test reports, and EL image metadata, to avoid cross-type matching failures. Strongly associated fields such as batch number and component ID need secondary indexes to enable fast associated recall, supporting cross-checks of the same batch of materials and module processes. The length of single data entries varies widely, from tens of characters for batch records to hundreds of characters for process parameters. Vector models must adapt to different input lengths to ensure complete feature extraction.
How to Set the Configuration
The following configurations are based on the vector model and index settings interface of FastGPT v4.17.0. If a free index model is needed, open-source vector models and reranking models can be selected for local deployment.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| maxToken | 8192–16384 token | The text length of photovoltaic module process records and EL image metadata mostly falls within the thousands of token range. Adapting to this range ensures complete feature extraction. |
| defaultToken | 4096 token | Short text batch records and packing lists do not need to occupy excessive token resources, ensuring basic matching efficiency. |
| Chunk Length | 800–1200 characters | Splitting process records and test reports for a single module can retain complete semantic units, avoiding loss of cross-chunk information. |
| Overlap Length | 100–150 characters | Prevents semantic breaks between adjacent chunks after splitting, ensuring consistent vector matching for cross-chunk associations. |
| Similarity Threshold | 0.72–0.80 | Subtle differences in cell current gears need to be distinguished. A threshold that is too high will miss valid matches, while one that is too low will introduce invalid associations. |
| Single Search Limit | Top 20 entries | The number of associated documents for the same batch of materials and modules typically falls within the 10-15 entry range. This covers all associated data while avoiding redundant calculations. |
Three Common Misconfigurations
- Phenomenon: Vector recall results do not include associated records for the same batch of materials, and the number of returned results is far lower than expected. Cause: A secondary index for the batch number field was not created, or the vector storage backend was not correctly configured with the target cluster address, preventing the index from quickly locating associated data for the same batch.
- Phenomenon: The interface returns a
413 Request Entity Too Largeerror. Cause: A reasonablemaxTokenparameter was not set. The input EL image metadata or long text process record exceeds the maximum token limit allowed by the model. - Phenomenon: Cross-review matches a large number of unrelated cell grading records, leading to a high false positive rate. Cause: A private reranking model was not integrated, or the similarity threshold was set too low, including weakly associated results that do not meet process matching requirements into the review scope.
How to Confirm the Configuration Is Correct
- Upload all associated documents for a single module, run a cross-review query, and verify that returned results include all batch material records and process parameters corresponding to that module.
- Review vector model call logs to confirm input text token counts stay within the set
maxTokenrange, with no truncation errors. - Adjust the similarity threshold, compare recall result counts across different thresholds, and confirm the threshold range meets the matching accuracy requirements of the current business.
- Check the connection status of the vector storage backend, and confirm that configurations migrated from PGSQL have been correctly updated to the target cluster address.
- Call the reranking model interface, and verify that returned sorted results align with the priority requirements of photovoltaic module process matching.
The values provided are common starting points and should be measured against the reader's own samples.