What this category of data looks like
Data for motor and generator manufacturing processes and standards comes from electromagnetic design databases, copper wire and insulation material batch ledgers, winding equipment logs, dynamic balancing machine records, electrical test bench reports, and ERP and MES systems. Daily generated documents include winding process cards, insulation treatment records, dynamic balancing reports, dielectric withstand voltage test sheets, type inspection reports, and factory inspection reports. Data is updated per individual motor or product batch, and synchronized after each type modification. Document fields include winding turns, insulation class, test voltage, balance accuracy value, and more. Units include turns, V, μm, g, and others. The structure splits into fixed parameter entries and process description sections.
Constraints on knowledge base retrieval and recall
This category of data has high update frequency and tight association between single-batch datasets. The retrieval system must support incremental synchronization and batch associated recall to avoid data lag or fragmented disconnection. Documents contain two types of content: fixed parameters and free text. The system must balance exact field matching and semantic understanding of descriptive content. Single document lengths vary widely, from dozens-of-word process card entries to thousands-of-word test reports. A chunking strategy that adapts to different content lengths is required. Retrieval result accuracy directly impacts production compliance and safety. The proportion of irrelevant mixed results must be strictly controlled.
How to configure the system
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
Similarity threshold (corresponding to similarity) | 0.72–0.78 | Matches fixed field accuracy for motor manufacturing documents, balances recall rate and accuracy |
Single search limit (corresponding to limit) | Top 8 entries | Covers reference data across winding, insulation, and test links, avoids excessive results interfering with judgment |
Rerank switch (corresponding to usingReRank) | Enabled | Improves sorting accuracy for long document fragments, prioritizes returning the most relevant process steps |
Search mode (corresponding to searchMode) | Keyword + semantic hybrid | Adapts to fixed fields like winding turns and insulation class on process cards, and descriptive analysis content in test reports |
Query optimization switch (corresponding to datasetSearchUsingExtensionQuery) | Enabled | Supplements synonymous expressions for process terms, such as matching "inter-turn insulation" to "winding insulation" |
Dataset ID (corresponding to collectionIds) | Exclusive motor manufacturing process library ID | Limits retrieval scope, excludes knowledge base content from other categories |
Three common configuration mistakes
- Symptom: Retrieval returns no results or irrelevant general machinery documents. Cause:
collectionIdsis not set to limit the exclusive dataset, resulting in mixed content from other category knowledge bases. - Symptom: Knowledge base training task throws an error, and the interface shows an abnormal training status. Cause: Incremental synchronization is not configured according to the per-unit or per-batch update rhythm, and single synchronization data volume exceeds the system processing limit.
- Symptom: Exact matching results for winding turns cannot be retrieved. Cause: The
similaritythreshold is set too high, ignoring reference content with similar turns on process cards, or thedatasetSearchUsingExtensionQueryswitch is not enabled to supplement term matching.
How to confirm the configuration is properly set
- Upload a test snippet of a winding process card, run a retrieval that includes "winding turns", and verify if returned results include relevant fragments of this document.
- Check the FastGPT v4.17.0 dataset synchronization log, confirm that process documents for each motor are retrievable within 1 hour after update.
- Adjust the
similaritythreshold, compare retrieval results across different thresholds, and confirm that the match between returned content and process questions meets business expectations. - After enabling
usingReRank, verify that sorted results place the most relevant process steps first.
The values provided are common starting points and should be measured against the reader's own samples.