This configuration is based on FastGPT v4.17.0, and is adapted to the unique data characteristics of this category.
What data for this category looks like
Data for this category comes from PLM, ERP, MES, supplier materials, validation tests, and customer-specific requirements. Updates are triggered irregularly alongside 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 files. Most documents contain fields such as change number, version number, effective date, associated materials or processes. No fixed uniform standard exists for single-document length.
Constraints on knowledge base retrieval and recall
Scattered data sources across multiple systems require retrieval to cover documents in heterogeneous formats. Irregular updates triggered by changes require retrieval indexes to support incremental synchronization, ensuring latest change documents are recalled in a timely manner. Documents contain unique fields including change number, version number, and effective date, so precise filtering by version and associated process must be supported. Wide variation in single-document length requires retrieval chunking to adapt to different content lengths, avoiding context window overflow. Changes associate with multiple materials and processes, so precise recall of documents within the associated scope is needed to prevent irrelevant results from interfering with Q&A accuracy.
How to configure the system
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Most documents in this category are structured change records, so low-relevance results must be filtered to avoid mixing irrelevant general process documents |
limit | Top 8–12 results | The number of documents associated with a single change is limited. Too many results will exceed the context window and cause token overflow |
usingReRank | Enabled | Multiple associated fields exist in documents of this category, and re-ranking improves matching accuracy for change scenarios |
datasetSearchUsingExtensionQuery | Enabled | Change-related queries often include keywords such as version and number. Optimization supplements associated retrieval terms and expands recall scope |
searchMode | fullTextRecall | The query examples emphasize identifiers or specialist terms; test this mode with representative data, not as a product default. |
collectionIds | Bind to the dedicated change management knowledge base collection | Avoid retrieving general equipment documents unrelated to changes, and narrow the retrieval scope |
Three common configuration mistakes
- Phenomenon: Using an existing general search engine template prevents change-related document searches from triggering normally. Cause: The template does not bind to the dedicated change management knowledge base collection for this category, so no available data sources exist for retrieval.
- Phenomenon: Setting
limitto 2000 results causes the single retrieval result set to exceed the context window, and the answer token count exceeds 3000. Cause: The recall upper limit is set too high without accounting for the limited number of change-associated documents for this category. - Phenomenon: After enabling retrieval, results are output directly on the interface, and subsequent AI replies repeatedly display retrieved content. Cause: The connection logic between the retrieval node and the AI node is not correctly configured in the workflow, causing retrieval results to be output repeatedly.
How to verify correct configuration
- Manually enter a test question containing a known change number, and confirm retrieval results only include change management documents for this category.
- Adjust the
similaritythreshold, verify retrieval result relevance meets expectations, and confirm the threshold fits the characteristics of documents in this category. - Check the connection logic between the retrieval node and the AI node in the workflow, confirm retrieval results are only passed to the AI node as context, and are not directly output to the interface.
- Trigger a test Q&A related to changes, confirm the token count within the context window does not exceed system limits, and verify the
limitsetting is reasonable.
The values provided are common starting points and should be measured against the reader's own samples.