What Data for This Category Looks Like
Daily documents for this category include grouping rules, BOMs, welding torque records, BMS flash records, insulation and voltage withstand reports, EOL test data, and system configuration lists. Data sources include cell capacity grading data, MES systems, welding equipment, torque tools, BMS software versions, and EOL test benches. Data updates occur once per produced power battery pack or energy storage cabinet. Updates are also triggered by each BMS software version change. Documents include numeric fields such as torque (unit: N·m) and cell capacity (unit: Ah), as well as text fields like version numbers and test results. Both structured tables and unstructured test reports are included.
What Constraints These Characteristics Impose on Knowledge Base Retrieval and Recall
Data sources for this category are scattered and updates are frequent. Knowledge base retrieval must support incremental synchronization mechanisms to avoid recalling outdated process data. Document types include structured production records and unstructured test reports. Retrieval must adapt to recall logic for multi-format data. Individual production documents contain specialized content and clear fields. Retrieval must accurately match process terminology and parameters to avoid irrelevant results. Incorrect process parameters can cause serious issues such as thermal management failure and BMS protection malfunction. Retrieval result confidence levels must meet production safety requirements, and the relevance threshold of recall results must be strictly controlled.
How to Configure Settings
This document is written based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | This category of process documents is highly specialized. Filtering low-relevance retrieval results ensures the accuracy of recalled content |
limit | Top 8 results | Individual production documents have lengthy content. Too many recall results will exceed the context window limit, affecting subsequent large model processing performance |
searchMode | mixedRecall | This category of data includes structured field data and unstructured test reports. Hybrid retrieval can adapt to both data formats |
usingReRank | Enabled, re-rank top 5 results | Semantic similarity judgments for specialized process terms can have errors. The re-ranking mechanism further improves the relevance of recall results |
datasetSearchUsingExtensionQuery | Enabled, extend query length to 200 characters | This category contains a large number of specialized process terms. Extended queries can cover synonyms and industry standard expressions, improving recall coverage |
collectionIds | Bind specified collection IDs for the corresponding product line | Process parameters vary significantly across different product lines. Binding specified collections avoids recalling content from irrelevant knowledge bases |
Three Common Misconfigurations
- Symptom: A 400 error is returned when passing the
datasetIdvariable in the chat API call. Cause: The variable is not configured as a knowledge base authorization type, or the variable format does not match the string format of the dataset ID. - Symptom: Exported knowledge base files contain outdated, unupdated data. Cause: Knowledge base incremental synchronization was not triggered, so modified CSV files or documents were not synchronized to the retrieval index.
- Symptom: Target documents cannot be located when querying a single PDF. Cause: The target PDF was not created as a separate collection, or the
collectionIdsparameter was not specified in the retrieval configuration to bind the corresponding collection.
How to Verify Successful Configuration
- Upload a test grouping rule document, initiate a query containing keywords from that document, and confirm returned results include the target document content.
- Review retrieval logs, and confirm the
similarityscore of recall results falls within the preset interval. - Check knowledge base update records, and confirm index synchronization status is completed after each production data change.
- Pass the specified
collectionIdsparameter, and confirm retrieval results only originate from the bound knowledge base collection.
The values provided are common starting points and should be measured against the reader's own samples.