What this category of data looks like
Data for this category is generated by a change management team from manufacturing enterprises, with participation from R&D, process, quality, production and procurement departments. Daily documents include ECN/ECR, change evaluation forms, drawings and BOM version files, process verification reports, production switching plans, inventory disposal documents, customer approval documents, and more. Data sources cover PLM, ERP, MES systems, supplier materials, verification test records and customer special requirement documents. Update frequency is irregular, triggered each time a process or engineering change occurs. Most documents include fields such as version number, effective date and associated material code.
Constraints imposed by these characteristics on the knowledge base retrieval and recall link
The multi-system sources, irregular updates, multi-type document structures and versioned fields of this category of data require the retrieval and recall link to support cross-data source association matching, and prioritize recalling the latest version of documents. Documents contain multi-format content such as drawings, BOMs and verification reports, so chunking and embedding rules for different file types must be adapted. The irregular updates triggered by changes require the retrieval link to support incremental synchronization mechanisms to avoid recalling outdated old-version files. At the same time, changes are associated with chain effects on materials, processes and inspections, so retrieval must accurately match associated fields to ensure recalled content covers all chain documents involved in the change.
Configuration settings
These configurations are formulated based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Documents in this category require matching precise fields such as version numbers and associated material codes. A threshold that is too low will introduce irrelevant change documents, while a threshold that is too high will miss valid associated content |
limit | Top 8–12 entries | Change review needs to cover full-chain associated documents. This range balances the number of recalled results and context length |
searchMode | mixedRecall | Documents in this category include structured fields and unstructured text. Hybrid retrieval can match both semantics and precise keywords |
usingReRank | Enabled | Documents recalled from multiple sources need to be sorted based on semantic relevance and version age. Reranking optimizes result priority |
datasetSearchUsingExtensionQuery | Enabled | Change requirements often contain implicit associations. Extended queries can supplement matching of associated documents that are not directly mentioned |
collectionIds | Specify dedicated collection IDs for change management | Avoid recalling irrelevant production scheduling and routine inventory documents, and narrow the retrieval scope |
Three common errors
- Phenomenon: Image-format ECN documents are not included in retrieval results. Cause: The image dataset support configuration is not enabled in the private deployment environment, so image-based change documents cannot be imported or recalled.
- Phenomenon: Historical call logs for individual documents cannot be viewed. Cause: The retrieval log collection function is not enabled, so call records are not retained by the system.
- Phenomenon: Retrieval tasks return an abnormal error containing the
$schemafield. Cause: The document metadata in the dataset has an invalid schema definition, causing the embedding process to fail to parse.
How to confirm the configuration is effective
- Initiate a retrieval request associated with a specific material code, check whether the recalled results include the ECN and process verification report of the corresponding change, and confirm that the retrieval scope matches the configuration.
- View the sorting of retrieval results, confirm that the latest version of documents is ranked first, and verify the role of the reranking and similarity configuration.
- Check the system logs, confirm that retrieval call records have been normally collected, and verify that the log-related configuration is effective.
- Import a compliant drawing document, confirm that it can be successfully indexed and recalled, and verify that the dataset configuration covers the corresponding file type.
The values given are common starting points and should be measured against the reader's own samples.