What Data for This Category Looks Like
Data sources include design BOMs, silicon steel sheet and insulating material ledgers, vacuum drying equipment operation logs, oil test reports, and partial discharge and withstand voltage test records from test stations. Data is updated per individual product or batch of raw materials. Full process documentation for a single product includes winding and core assembly records, vacuum drying records, insulating oil test reports, partial discharge withstand voltage test reports, and nameplate parameter tables. Documentation includes dedicated fields such as insulation moisture content, test voltage, and partial discharge values, paired with industry-specific units including ppm, kV, and pC.
Constraints Imposed by These Characteristics on Vector Models and Indexing
Full process data for a single product has strong internal correlation. Indexes must support filtering and recall by batch, serial number, and other dimensions to prevent interference from cross-product data. The high-frequency single-update feature requires indexes to support incremental update logic, avoiding the time costs of full reconstruction. The mixed multi-type document structure requires indexes to support mixed embedding of structured fields and semi-structured text. High-precision numerical fields such as insulation parameters and test values raise higher requirements for the semantic alignment accuracy of vector embeddings. For scenarios with small-batch, high-frequency writes, balance must be struck between index construction efficiency and recall accuracy.
Configuration Settings
This document applies to FastGPT v4.17.0. The following configurations are based on this version:
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the paragraph length of single product documentation, avoiding splitting of cross-process information |
| Overlap Length | 100–150 characters | Retains cross-chunk process association information, preventing critical test data from being truncated |
| Similarity Threshold | 0.75–0.85 | Matches the semantic matching accuracy of insulation parameters and test values, filtering irrelevant recall results |
| Single Search Limit | Top 10 entries | Matches the volume of associated data for a single product, avoiding interference from redundant information |
maxToken | 4096 | Matches the long text content of transformer test reports, fully loading critical parameters |
defaultToken | 2048 | Matches the text length of conventional assembly records, balancing recall and inference costs |
Three Common Mistakes
- Phenomenon: Vector index construction shows no progress for an extended period, with no clear background error. Cause: Incremental indexing is not configured for the single-product high-frequency update business characteristic, and full reconstruction logic is accidentally triggered.
- Phenomenon: Searches return data from other product batches unrelated to the current inspection question. Cause: No batch or serial number filtering rules are set for a single product, or the similarity threshold is set too low.
- Phenomenon: Embedded insulation parameters and test values have insufficient matching accuracy. Cause: A general text embedding model is used to process structured numerical fields, without adapting to the parameter semantics specific to this category.
How to Confirm Proper Configuration
- Upload the full documentation set for a single product, confirm that index construction progress matches expectations and no abnormal error messages appear.
- Input an inspection question targeted at that single product, verify that recall results only include relevant documentation from the same batch or model.
- Review embedded field details, confirm that structured fields such as insulation parameters and test values have been correctly converted to vector format.
- Add a new inspection record for a single product, test whether the index automatically synchronizes the new content without requiring manual full reconstruction.
The values provided here are common starting points, and should be evaluated against the reader's own samples.