What this category of data looks like
Data for this category originates from manufacturing enterprise PLM, ERP, MES, supplier materials, validation tests, and customer-specific requirements. Updates trigger irregularly with each process or engineering change. Daily documents include ECN/ECR forms, change assessment reports, drawings and BOM version records, process validation reports, production switching plans, inventory disposal documents, and customer approval documents. Documents contain fields such as version number, change item, effective time, associated materials or processes. No unified fixed unit exists; some drawing documents include dimensional tolerance units.
What constraints these characteristics impose on knowledge base retrieval and recall
Data sources are scattered and connected to multiple systems, requiring cross-source retrieval of associated change documents across systems. Irregular updates triggered by changes require retrieval to match the latest versions in real time, avoiding use of outdated invalid data. Document structures vary widely, including both structured forms and unstructured drawing text, requiring adaptation to different segmentation and retrieval rules. The version field of change-related documents is a core retrieval condition, requiring precise matching of version numbers to avoid confusion between new and old versions. A single change involves a large number of associated documents, requiring control of the recall scope to avoid returning excessive redundant content.
How to configure settings
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Matches precise text related to changes, avoiding recall of irrelevant historical documents |
limit | Top 8–12 entries | Controls the number of associated documents recalled per request, adapting to the multi-document association scenario of change audits |
searchMode | mixedRecall | Covers both structured fields such as version numbers and unstructured text content |
usingReRank | Enabled | Rearranges recall results by change relevance, prioritizing high-priority documents such as validation reports and customer approval documents |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands change-related keywords based on user questions to supplement retrieval terms |
collectionIds | Specify dedicated change management dataset IDs | Limits the retrieval scope, excluding non-associated documents from other categories |
Three common mistakes
- Phenomenon: After uploading a multi-column Excel format change material list, the automatically segmented results displayed on the interface are chaotic, with each row of material information split into multiple text blocks. Cause: The segmentation rules are not adjusted for structured table data, and the default automatic segmentation does not split content by row boundaries.
- Phenomenon: Deleted old version change documents marked in the local knowledge base still display reference links in retrieval results. Cause: Incremental synchronization cleanup rules for the dataset are not configured, and old data is not completely removed from the vector database.
- Phenomenon: In hybrid retrieval mode, general process documents not associated with the current change are incorrectly recalled, interfering with audit judgments. Cause: The retrieval scope is not limited via
collectionIds, or thesimilaritythreshold is set too low, leading to excessively high matching scores for irrelevant documents.
How to confirm the configuration is correct
- Upload one typical ECN document, enter the retrieval question "What BOM versions are involved in this change", and check whether the recall results include associated BOM version records.
- Adjust the
similarityparameter range to verify that the relevance of recall results changes as expected with the parameter. - After enabling
usingReRank, confirm that high-priority documents such as customer approval documents and process validation reports appear at the top of the retrieval results. - Check that the retrieval scope only covers the dedicated change management dataset, confirming that documents from non-associated categories are not recalled.
The values provided are common starting points and should be evaluated against the reader's own samples.