What data for this category includes
Data for this category comes from winding, insulation, final assembly, and testing departments of electric motor and generator manufacturers. Daily documents include winding process cards, insulation treatment records, dynamic balance reports, voltage withstand insulation test records, and type and factory acceptance test reports. Data sources cover electromagnetic design data, copper wire and insulation material batch information, winding equipment, balancing machine data, electrical test bench data, and ERP/MES system data. Data is updated upon completion of each product or production batch. All associated documents are synchronized after each product type change. Documents contain clear fields for process parameters, test values, batch identifiers, and change records. Units include category-specific metrics such as turns, temperature, pressure, and power.
Constraints on vector models and indexing
The high-frequency small-batch updates, mixed multi-type test documents, and dedicated process parameter fields of this category’s data impose multiple constraints on the vector model and indexing workflow. First, documents for each product or batch are independently linked to batch information. Chunking must retain contextual associations between batch identifiers and process nodes to avoid cross-batch confusion of process requirements during retrieval. Second, documents contain a large number of numerical parameters with dedicated units. Vector models must adapt to mixed structured and unstructured text content to ensure accurate matching of process parameters and abnormal scenarios during semantic retrieval. Additionally, the high-frequency update feature requires indexes to support incremental synchronization, avoiding time-consuming issues caused by full index reconstruction. It also ensures retrieval results stay consistent with the latest process changes and test data.
Configuration settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Matches the single-paragraph length of winding process cards and test reports, preventing semantic fragmentation of critical process steps |
| Overlap Length | 100–150 characters | Retains contextual information for cross-chunk winding wiring and insulation treatment workflows, preventing key information from being truncated |
| Similarity Threshold | 0.72–0.85, calibrated via actual testing | Matches the semantic matching accuracy required for this category’s process parameters and abnormal scenarios, avoiding mismatched process requirements across different batches |
| Single Search Limit | Top 6 results | Controls the volume of retrieval results, aligning with single-scenario query needs for process Q&A |
maxToken | 4096 token | Aligns with the model input limit of FastGPT v4.17.0, ensuring full loading of associated process documents and test data |
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 configuration errors
- Issue: After the vector index is created, the PG database only stores vector data and cannot associate batch and process parameter content from the original documents. Cause: The original document persistence configuration item was not enabled, and only the vectorized results were synchronized. This prevents retrieval from recalling complete information from the original process documents.
- Issue: When using the bge-m3 vector model, semantic retrieval returns abnormally high similarity scores, or the search test fails directly. Cause: The similarity threshold was not adjusted for this category’s dedicated numerical parameters, or the chunk length exceeds the model’s supported range, causing the semantic matching logic to fail.
- Issue: In a deployed FastGPT instance, the knowledge base indexing step remains unresponsive for an extended period, with task progress stuck. Cause: Batch synchronization indexing rules were not configured. When importing multi-batch motor test reports in full, the single task occupies too many system resources, leading to a timeout.
How to verify correct configuration
- Log in to the associated PG database, view the data table corresponding to the knowledge base, and confirm that both original document text fields and vector fields exist. This verifies that the storage configuration is active.
- Input queries for dedicated scenarios such as winding turns verification or insulation treatment step inquiries, and confirm that the similarity scores of retrieval results fall within the preset threshold range. This verifies that parameter configurations align with requirements.
- Import a single batch of insulation treatment record documents, perform incremental index synchronization, and confirm that the task completes within a reasonable time frame. This verifies that the update configuration adapts to high-frequency small-batch update needs.
- Call the FastGPT retrieval interface, check that the number of returned results matches the single search limit setting. This verifies that the retrieval rules are active.
The values provided are common starting points and should be measured against the reader's own samples.