What the data for this category looks like
The core data for this category comes from manufacturing enterprise IPQC, FQC, quality engineering and laboratory teams. Daily documents include control plans, patrol inspection records, SPC reports, final product inspection reports, nonconforming product review forms, release forms and quality monthly reports. Data sources cover MES systems, test equipment outputs, laboratory records, process parameters, customer standards and historical defect databases. Updates are collected continuously per batch and per shift. Documents are stored classified by production batch and inspection link, with fields including batch identification, inspection parameter thresholds, defect types, release judgment standards and more. There is no unified fixed format, but all are associated with specific production units.
What constraints these characteristics impose on the knowledge base retrieval and recall link
This category’s data is updated frequently per production batch. The retrieval pipeline must support filtering recall results by batch dimension to avoid invalid cross-batch content interfering with judgments. Multiple document types are stored together, so retrieval must adapt to different matching rules for structured reports and semi-structured review records, balancing semantic and field accuracy. Data sources include customer standards and internal process parameters, so the recall range must be limited to compliant authorized documents to prevent unauthorized content from being called. Continuously collected real-time data requires regular incremental index synchronization to avoid stale data being prioritized for recall, which affects inspection judgments.
How to set the configurations
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.72–0.82 | Matches the accuracy requirements for in-process quality documents, balances recall coverage and noise filtering |
limit | Top 6–10 entries | Adapts to the volume of documents related to single-batch inspections, avoids context overload that reduces audit efficiency |
searchMode | embedding | Adapts to the semantic and field matching requirements for structured inspection reports and control plans |
usingReRank | Enabled | Performs secondary ranking on recall results, improves the relevance of content related to nonconforming product reviews and release standards |
datasetSearchUsingExtensionQuery | Enabled | Expands question semantics to cover synonymous expressions for process parameters and defect types |
collectionIds | Configured by production support knowledge base groups | Limits the recall range to only document collections related to in-process quality and final product release |
Three common errors
- Phenomenon: Retrieval result citations appear only at the end of a response, not embedded in contextually relevant positions. Cause: No retrieval result context splicing rule is configured, only the original citation list is retained.
- Phenomenon: Retrieval results do not include the latest per-shift patrol inspection records. Cause: Incremental synchronization updates are not configured, so the knowledge base index has not synced real-time data from MES systems and test equipment in a timely manner.
- Phenomenon: The large language model does not reference knowledge base content when running in non-tool call mode. Cause: Automatic retrieval and recall association configuration is not enabled, or the target knowledge base collection is not bound.
How to confirm the configuration is correct
- Upload a latest control plan document, run a query that includes the inspection thresholds from this document, and verify that the recall results include this document.
- View the FastGPT v4.17.0 knowledge base index logs, confirm that per-shift and per-batch update data has completed synchronization with no failed records.
- Switch between different document collection groups, verify that the recall range configured by
collectionIdstakes effect, and only returns documents from the target category. - Run a query that uses synonymous process parameters, confirm that relevant content is correctly recalled after enabling the
datasetSearchUsingExtensionQueryconfiguration.
The values given are common starting points and should be measured against the reader's own samples.