What the Data for This Category Looks Like
Data for this category comes from electromagnetic design documents, copper wire and insulating material batch records, winding equipment logs, balancing machine operation data, electrical test bench test results, and production and material data from ERP/MES systems. Daily generated documents include winding process cards, insulation treatment records, dynamic balance reports, voltage withstand insulation test reports, and type and factory inspection reports. Data is updated upon completion of each product or batch production, and associated documents are updated synchronously when the product type is changed. Document fields include winding turns, paint dipping and drying parameters, vibration amplitude, temperature rise value, voltage withstand level, etc., with clear process parameters and test units.
Constraints on Knowledge Base Retrieval and Recall
Data for this category is updated per single product or batch, and contains a large number of professional process and test parameters, which imposes multiple constraints on the knowledge base retrieval and recall link. First, retrieval must associate identifiers such as production batch and equipment number to accurately locate the process and test data of the corresponding product. Second, professional fields such as winding turns and voltage withstand level require precise matching to avoid irrelevant results interfering with analysis. Data formats from multiple sources vary, so unified chunking rules are needed to ensure consistent recall. Frequently updated production and test data require index update rhythm to match production tempo to prevent data lag. At the same time, defect analysis needs to recall both process parameters and test results, and the content scope of a single recall must cover associated document types.
Configuration Setup
This configuration is based on FastGPT v4.17.0, with specific configuration items and basis as follows:
| Configuration Item | Recommended Setting | Basis for This Setting |
|---|---|---|
similarity | 0.75–0.85 | This category requires precise matching of professional parameters such as winding turns and voltage withstand level to avoid low-relevance results interfering with defect analysis |
limit | Top 10 entries | A single recall must cover multiple document types such as process cards and test reports for the same batch, while controlling context length |
| Chunk length | 800–1200 characters | Matches the paragraph length of documents such as winding process cards and test reports to avoid splitting of key parameters |
| Chunk overlap length | 120 characters | Ensures professional content spanning paragraphs such as winding turns and paint dipping parameters can be fully recalled |
usingReRank | Enabled | Semantic similarity judgment for professional data is easily disturbed by parameter units, and re-ranking can improve result relevance |
datasetSearchUsingExtensionQuery | Enabled | Can expand synonyms related to processes and tests to improve recall rate of professional content |
collectionIds | Bound to exclusive document collection ID for motor manufacturing | Only recalls process, test and production data for this category to avoid mixing irrelevant knowledge base content |
Three Common Errors
- Symptom: After uploading documents, the index status of the knowledge base dataset stays at "Indexing" for a long time with no progress updates. Cause: The rhythm of frequently updated production data for this category is not matched, the set index trigger frequency does not match the production update frequency, or the number of documents imported in a single batch exceeds the system's single processing limit.
- Symptom: After replacing the embedding model, the similarity score of retrieval results reaches 10000+, and irrelevant results cannot be filtered by threshold. Cause: The
similaritythreshold is not recalibrated. The similarity score range of the new model is inconsistent with that of the original model, causing the threshold setting to fail. - Symptom: After uploading documents, there are no corresponding index entries in the knowledge base, and the background log shows no available channels. Cause: The available channel for the corresponding embedding model is not configured in the system, causing documents to fail vectorization processing.
How to Verify Successful Configuration
- Upload a single winding process card of this category, wait for indexing to complete, retrieve preset professional keywords, and check whether the relevance of the recall results meets analysis needs.
- Retrieve the dynamic balance report of products from the same batch, check whether the recall results include all associated document types for this batch, and whether the quantity matches the single search upper limit setting.
- After replacing the embedding model, adjust the similarity threshold again, and verify that the threshold can effectively filter irrelevant results by retrieving known correct process parameters.
- View the index log in the knowledge base backend, confirm that documents of this category have completed vectorization processing with no error messages.
The values given are common starting points and should be measured against the reader's own samples.