What the data for this category looks like
This category’s core data typically comes from PLM/ERP systems, parts inspection databases, assembly tool collection terminals, test benches, MES systems, and customer technical agreements. Daily documents include assembly process cards, multi-level BOMs, torque records, gap adjustment records, trial run reports, factory inspection sheets, and packing lists. Data updates follow the completion rhythm of each equipment assembly order. Each device has exclusive assembly, inspection, and delivery documents. Documents contain fields with clear units, such as newton meters, millimeters, and megapascals, including device model, component number, torque value, gap parameter, and test result.
Constraints Imposed on Knowledge Base Retrieval and Recall
The multi-level BOM structure of this category requires the retrieval link to support cross-component and cross-step associated recall, to meet the need of verifying physical parts against documents during assembly. The exclusive update rhythm per single device requires the knowledge base to support incremental recall by order or device ID, avoiding redundant overhead from full retrieval. Fields with clear units such as torque values and gap parameters require unit association during retrieval, to prevent incorrect matching of ununitized parameters. Format differences across heterogeneous data from multiple sources require the retrieval link to support unified parsing and field alignment across PLM, MES and other systems, improving recall accuracy.
Configuration Settings
The following configurations are based on the parameter logic of FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Documents for this category mostly contain precise parameters. Low-match irrelevant content must be filtered to retain highly relevant assembly and test data |
limit | Top 10 entries | Assembly scenarios require quick access to core process steps and fault handling solutions. Too many results increase screening costs |
searchMode | mixedRecall | This category has both structured parameter fields and natural language descriptions of assembly steps. Hybrid mode balances precise matching and semantic recall |
usingReRank | Enabled | Hybrid retrieval results may have sorting deviations. Reranking re-sorts based on semantic relevance, improving the priority of core results |
datasetSearchUsingExtensionQuery | Enabled | Engineer questions often omit prefixed information such as component numbers. Optimization supplements associated fields, improving recall accuracy |
collectionIds | Bind PLM and MES document collections corresponding to the order | Data for this category updates per order. Binding exclusive collections avoids interference from cross-order data, improving single-device retrieval accuracy |
Three Common Misconfigurations
- Phenomenon: After enabling
usingReRank, no results are returned in the application's retrieval, but the knowledge base backend test returns results normally. Cause: The reranking model's call permissions are not linked to the current application, or the reranking model's input parameter format does not match the application layer configuration, causing the call to interrupt. - Phenomenon: Knowledge base retrieval takes too long, with a single query exceeding the preset threshold. Cause: No exclusive knowledge base collection is bound by order or device ID, and full retrieval of multi-level BOMs and assembly documents across orders causes excessive recall overhead.
- Phenomenon: Retrieval results do not include the current device's assembly process card content. Cause: No incremental update logic is configured, old versions of assembly documents are not synchronized to the current knowledge base collection, or the collection ID is configured incorrectly, pointing to an unrelated document library.
How to Confirm Proper Configuration
- Enter the knowledge base test page in the FastGPT backend, enter a search term containing the device model and component number, and check whether the returned results include the assembly process card and BOM content for the corresponding order.
- After enabling
usingReRank, initiate a retrieval in the application, compare the sorting of backend test results and the sorting displayed in the application to confirm that the reranking function is active. - View the knowledge base update log, confirm that incremental data generated per order has been synchronized to the bound collection, with no missing assembly or inspection documents.
- After adjusting
similarity, verify in the backend test whether low-match irrelevant content is filtered, confirming that the threshold configuration meets the current scenario requirements.
The values given are common starting points and should be measured against the reader's own samples.