What the data for this category looks like
This category’s daily documents cover winding and core assembly records, vacuum drying records, insulating oil reports, partial discharge and withstand voltage test reports, and nameplate parameter tables. Data sources include design BOMs, silicon steel sheet and insulating material procurement data, vacuum and drying equipment logs, oil test reports, and test station test records. Updates to data occur once per product or per batch of materials. Each individual document includes clear professional parameter fields such as insulation distance, water content, and test voltage value. Common units include millimeters, percentage, kilovolt, and the document must tie to the traceability information of a single product.
What constraints do these characteristics impose on the vector model and indexing link
The data characteristics of this category impose multiple constraints on the vector model and indexing link. Each single product binds to exclusive traceability information. The indexing system must support precise associated retrieval by batch and serial number, preventing confusion of cross-product data. Professional parameter fields account for a high proportion. The vector model must adapt to mixed encoding of structured parameters and long-text reports, ensuring semantic consistency of professional information. Updates to data occur per product or per batch of materials. The indexing system must support incremental update logic, avoiding resource consumption caused by full reconstruction. The tender scenario requires association with historical quotation and test data. The indexing system must support cross-document associated recall, ensuring parameter accuracy during quotation preparation.
How to set the configurations
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the long-text structure of insulating oil reports and test reports, retains complete parameter context |
| Overlap Length | 100–150 characters | Avoids splitting continuous professional parameters such as insulation distance and partial discharge threshold |
| Similarity Threshold | 0.75–0.85 | Filters low-match historical test data to ensure parameter accuracy of tender documents |
| Reranking | Enabled, top 5 results | Filters the most relevant test reports and quotation data, adapts to the precise retrieval needs of tender scenarios |
| maxToken | 8192 | Adapts to the maximum input length of the Qwen3-Embedding-8B model, ensures complete encoding of long documents |
| defaultToken | Set from representative sample documents | defaultToken is the default token count for direct text chunking; tune it on representative source documents. See the v4.17.2 model-field reference. It is not a vector dimension, embedding-model input cap, or retrieval/LLM context. |
Three common mistakes
- When deploying Milvus, the service fails to start because the compose file includes PostgreSQL-related configurations. The cause is using a general deployment template with multiple database dependencies, without adjusting configurations for the lightweight retrieval needs of transformer manufacturing scenarios.
- When FastGPT connects to the Qwen3-Embedding-8B model deployed via VLLM, a connection exception error occurs. The cause is failing to correctly configure the model's API access address and port, or failing to match the model's input and output format requirements.
- The split documents and vectorization results of the knowledge base cannot be retrieved by product batch. The cause is failing to configure index binding rules based on the traceability information of a single product, leading to confusion of cross-product data.
How to confirm configurations are correctly set
- Upload an insulating oil report document. Check in FastGPT's document preview to confirm that the split chunks contain complete insulating oil parameters and test conclusions, with no obvious content fragmentation.
- Initiate a retrieval based on historical partial discharge test data. Confirm that the matching degree of returned results falls within the preset threshold range, with no irrelevant low-match data.
- Test the incremental update function. Upload a new vacuum drying record. Confirm that the index only updates the newly added data, and no full reconstruction operation is triggered.
- Check the vector database connection logs. Confirm that there are no timeout or authentication failure records, and the model call process operates normally.
The values listed here are common starting points, and should be evaluated against the reader's own samples.