What the data for this category looks like
The core data for this category comes from manufacturing enterprise PLM, ERP, and MES systems, as well as supplier materials, validation test documents, and customer special requirement documents. Daily generated documents include ECN/ECR forms, change assessment reports, drawings and BOM version files, process validation reports, transition plans, inventory disposal lists, and customer approval documents. Fields include version numbers, change nodes, material codes, process parameters, approval status, and more. Updates are triggered irregularly with each change, with no fixed cycle.
What constraints these characteristics impose on knowledge base retrieval and recall
Heterogeneous data from multiple sources (structured BOMs, semi-structured process records, unstructured drawings and PDF reports) requires retrieval that supports content parsing for different formats. Irregular change-triggered updates require support for incremental synchronization and real-time recall. Documents contain professional fields such as tables, long text, and version identifiers, requiring retrieval semantic matching accuracy aligned with industry terminology. A single change involves multiple types of related documents, requiring precise recall of all relevant items to avoid missing documents with chain effects. Cross-data-source retrieval for customer special requirements and supplier materials requires establishing dataset associations across different systems.
How to configure the settings
In FastGPT v4.17.0, the following configuration items can be used to adapt the retrieval requirements for this category:
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Matches semantic consistency for professional fields such as process parameters and version numbers, filtering low-relevance non-change documents |
limit | Top 8 entries | Covers multiple types of documents involved in a single change (BOMs, drawings, validation reports) while avoiding redundant recall content |
searchMode | mixedRecall | Combines keyword matching and semantic retrieval to adapt to mixed queries of material codes and process descriptions in ECN forms |
usingReRank | Enabled | Re-ranks recall results by change relevance, prioritizing returning version documents and validation reports directly related to the current change |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands query terms based on the current change topic, covering associated data sources such as customer special requirements and supplier materials |
collectionIds | Specify the dedicated dataset ID for change management | Restrict the retrieval scope to only change management-related documents, avoiding mixing in non-relevant data from other categories |
Three common configuration errors
- Uploaded ECN and BOM tables cannot be fully chunked and recalled, with material codes and process parameters in the tables missing from retrieval results. This occurs because no applicable chunking rules are configured for table documents, leading to truncated or incorrectly split table content.
- A "no knowledge base selected" prompt appears during retrieval, and no change-related documents are returned. This occurs because the
collectionIdsof the corresponding change management dataset are not specified in the retrieval entry, or target documents are not bound to the correct dataset. - Old and new version documents are mixed in recall results, with the latest validation reports and transition plans not prioritized. This occurs because
usingReRankis not enabled, or thesimilaritythreshold is set too low to distinguish version relevance.
How to verify the configuration is correct
- Upload an ECN table and a BOM version document, perform a retrieval based on the current change topic, and check whether the table content is fully recalled and chunked reasonably.
- Specify the
collectionIdsof the dedicated change management dataset in the retrieval interface, and confirm that the "no knowledge base selected" prompt does not appear during retrieval. - Compare the order of recall results when
usingReRankis enabled and disabled, and confirm that version documents and process validation reports are displayed first. - Adjust the
similarityparameter range, and check whether low-relevance non-change documents are effectively filtered.
The values provided are common starting points and should be evaluated against the reader's own samples.