What This Category’s Data Looks Like
The typical entities for this category are manufacturers of battery packs, energy storage cabinets, and BMS systems. Daily documents include grouping rules, BOM lists, welding and torque records, BMS flashing records, insulation and voltage withstand reports, EOL test data, and system configuration lists. Data sources include cell grading systems, MES production management systems, welding equipment, torque tools, BMS software, and EOL test benches. One record is generated for each produced battery pack or energy storage cabinet. Configuration-related documents are updated synchronously each time the BMS software version changes. Document fields include cell parameters, torque values, test results, software version numbers, and some fields have clear units.
Constraints on Knowledge Base Retrieval and Recall
Data sources for this category are scattered, covering multiple types of terminals such as production equipment, management systems, and testing tools. Retrieval must adapt to multi-source heterogeneous data formats. For example, original records exported from equipment and reports generated by systems need unified retrieval logic. Records for a single batch of products are updated frequently. Retrieval must support fast matching of newly generated grouping, testing, and flashing data to avoid returning outdated information. Daily documents include multiple types such as rule-based, record-based, and report-based documents. Retrieval must adapt to the search needs of different fields. For example, torque records need to match precise numerical ranges, and grouping rules need to be associated with cell parameter fields. Every product’s records are unique. Retrieval results must accurately correspond to the inspection and audit needs of the target batch to avoid mixing data across batches.
How to Configure Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Used to match numeric fields such as cell parameters and torque values. This interval balances retrieval precision and recall rate, and meets the strict requirements of inspection scenarios |
limit | Top 8 entries | The number of associated documents for a single batch of products is usually 5-7. Taking the top 8 entries covers all necessary data and avoids redundancy |
searchMode | fullTextRecall | The query examples emphasize identifiers or specialist terms; test this mode with representative data, not as a product default. |
usingReRank | Enabled | Re-ranks recalled results by relevance to inspection tasks, optimizing retrieval results for long texts such as welding records and EOL test reports |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonyms related to inspection. For example, expanding "insulation test" to "insulation voltage withstand test" to cover wording differences across documents |
collectionIds | Knowledge base collection bound to production batch | Ensures retrieval only accesses the document collection corresponding to the current batch, avoiding interference from cross-batch data |
Three Common Misconfigurations
- Phenomenon: Knowledge base search takes more than 12 seconds, and there is no obvious optimization for single-batch document retrieval. Cause:
usingReRankis not configured, or thesimilarityinterval is set too wide, resulting in an overly large retrieval range and excessive redundant data, which increases sorting and organizing time. - Phenomenon: After FastGPT is deployed offline, knowledge base query results are not organized according to business logic, and a large amount of irrelevant content appears. Cause:
collectionIdsis not configured for the locally deployed knowledge base collection, anddatasetSearchUsingExtensionQueryis not enabled, so it cannot adapt to the proprietary terminology of local documents. Ensure that FastGPT v4.17.0 or higher is used to ensure compatibility of parameter configurations. - Phenomenon: Retrieval results include BMS flashing records from other batches. Cause: The
collectionIdsbound to the corresponding production batch are not correctly configured, so the retrieval scope covers the entire knowledge base collection, leading to cross-batch data mixing and affecting the accuracy of inspection and audit.
How to Verify Configuration
- Retrieve the grouping rule document of a known batch, check whether the returned results only include the grouping parameters of the current batch, confirm whether the binding of
collectionIdstakes effect, and ensure that FastGPT v4.17.0 or higher is used. - Enter a query containing an exact numerical value, such as "torque value 25N·m", check whether the returned welding records match this value, and confirm whether the configuration of
searchModeandsimilarityis reasonable. - Test retrieval in an offline deployment scenario, check whether the returned results can be organized and output according to business logic, and confirm whether the configuration of
usingReRankanddatasetSearchUsingExtensionQuerytakes effect. - Retrieve a query with no corresponding content, such as "ungenerated cell grading parameters", check whether the returned results conform to the preset no-content processing logic, and confirm whether the retrieval boundary is correct.
The values provided are common starting points and should be measured against the reader's own samples.