What this category of data looks like
The core data for this category comes from PLM, ERP, MES, supplier materials, validation tests, and customer-specific requirements. Updates are triggered irregularly each time a change occurs. Document types include ECN/ECR forms, change assessment records, drawings and BOM version files, process validation reports, production switching plans, inventory disposal documents, and customer approval documents. These documents have dedicated fields such as change order numbers, version identifiers, associated process and material information. There is no unified fixed format template. Some documents include long text descriptions and structured tables. The size of a single data entry can reach several megabytes.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
Because data sources cover multiple systems and updates are triggered irregularly with each change, knowledge base retrieval must support filtering recall results by change batch or version dimension. This prevents mixing old and new versions which could cause production accidents. Documents include multiple types and structures. The retrieval system must adapt text chunk splitting for different lengths and formats. This ensures recalled fragments fully cover key information linked to changes. Changes involve associated data for cross-department collaboration. Retrieval must accurately match dedicated fields such as change order numbers and version identifiers. This narrows the recall scope and improves accuracy. At the same time, recall content must be strictly limited to official documents. This avoids introducing unofficial and uncertain information.
Configuration settings
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Change documents require high matching accuracy to prevent unrelated change records from being included in recall results |
limit | Top 8–12 entries | The number of valid documents linked to a change is limited. Too many recall results will add context redundancy |
searchMode | mixedRecall | Change documents include structured fields and text descriptions. The hybrid mode balances precise matching and semantic relevance |
usingReRank | Enabled | The relevance of change documents varies significantly. Reranking corrects sorting biases in initial recall results |
datasetSearchUsingExtensionQuery | Disabled | Change Q&A must strictly match official documents in the knowledge base. Avoid generating unofficial content |
collectionIds | Fill in the knowledge base folder ID bound to the change project | Accurately limit the recall scope. Prevent interference from cross-project or older version documents |
Three common configuration mistakes
- Phenomenon: Unable to filter knowledge base content in a specified folder during retrieval. Results include change documents from unrelated projects. Cause: The
collectionIdsparameter was not configured correctly. The knowledge base folder ID for the corresponding change project was not bound. - Phenomenon: Returned answers are not the original responses from knowledge base Q&A pairs. AI-generated supplementary content is present. Cause: The
datasetSearchUsingExtensionQueryparameter was enabled. This triggered additional query expansion, deviating from the requirement to strictly match official documents. - Phenomenon: Document fragments are lost after setting a 3000-character text chunk split. Recalled results are incomplete. Cause: The long text structure of change documents was not adapted. The splitting parameters did not match the paragraph logic of ECR/ECR forms. This caused long texts to be incorrectly truncated.
How to confirm the configuration is correct
- Run a test retrieval. Enter a keyword linked to a specific change. Check that recall results only include documents associated with that change. No cross-project content is present.
- Enter the original text of a preset knowledge base Q&A pair. Check that the returned result exactly matches the original response. No additional generated content is included.
- Review the vectorization task logs. Confirm that the embedding speed is within a reasonable range. No resource limit error messages are present.
- Check the split text chunks. Confirm that long documents have no missing fragments. Full change-related information is covered.
The values provided are common starting points. They should be evaluated against the samples relevant to the actual deployment scenario.