What the data for this use case looks like
Official marketing materials from business sub-brands under the group make up the primary source of marketing content data, including insurance product promotional scripts, financial plan manuals, offline event copy, and similar assets. The update rhythm aligns with the marketing activity cycles of each sub-brand. Bulk updates trigger when new products launch, events start, or compliance revisions are made, with partial content tweaks on a daily basis. Documents store and categorize content by sub-brand, business category, and content type, with fields including content ID, sub-brand identifier, business attribute, release time, content body, compliance check status, and more. The content body saves as plain text or rich text, with length measured in characters. Time fields use standard ISO format.
Constraints on Knowledge Base Retrieval and Recall
Different sub-brands store marketing content separately, so retrieval processes must isolate content by sub-brand to avoid recalling non-target business content across categories. Configure incremental update trigger rules for irregular bulk updates, to prevent full updates from consuming excessive system resources. The multi-field document structure requires setting field-level filter conditions to prioritize recalling content that passes compliance checks and matches business attributes. Set reasonable chunking rules for long-form marketing content, to avoid semantic cutting that destroys script integrity, while controlling the number of recalled items to fit context window limits.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
Recall Count | Top 8-12 items | Marketing content categories are scattered and total volume is large; too many recalled items increases context processing load, too few fails to cover all relevant content |
Similarity Threshold | 0.72-0.85 | Marketing scripts are mostly colloquial; a threshold that is too low may retrieve irrelevant content, while a threshold that is too high will filter out some marketing materials that meet the required matching degree |
Rerank Return Count | Top 3-5 items | Marketing content needs to precisely match user inquiry scenarios; retaining the most relevant content after reranking avoids excessive redundant information interfering with subsequent generation |
Incremental Update Trigger Threshold | 100 changes per batch | Adapts to the daily update scale of business sub-brands, avoids frequent full updates, and reduces incremental synchronization delay |
Field Filter Rules | Follow Compliance Status = Compliant and Sub-brand = Specified Value | Isolates marketing content from different sub-brands, ensuring recall results meet regulatory requirements and business scenario limits |
Chunk Length | 800-1200 characters | Adapts to the semantic integrity of long marketing text; cutting too short destroys script logic, while cutting too long increases computational load for single-chunk retrieval |
The parameter values provided on this page are common starting points for configuration. Actual values are affected by material format, data volume, and business rules. Specific issues require individual analysis, and it is recommended to test on your own samples before finalizing settings.
Three Common Misconfigurations
- Phenomenon: The rerank result field returns
falseafter calling the knowledge base retrieval node. Cause: The output switch for rerank model results is not enabled, or the rerank model is not bound to the current retrieval workflow. - Phenomenon: No optional content appears in the reference variable dropdown after selecting variable application in the "Knowledge Base Search" node. Cause: A variable of type "Knowledge Base ID" was not created in global variables, or the variable is not configured to be referenceable by workflows.
- Phenomenon: Retrieved content does not match user questions as expected. Cause: The similarity threshold setting is not adapted to the colloquial nature of marketing scripts, or field filter rules are not configured correctly.
How to Confirm Configuration is Complete
- Enter the knowledge base management page, check that field filter rules for compliance status and sub-brand have been added, confirm the configuration matches the business scenario.
- Trigger the knowledge base retrieval node in workflow debug mode, view the number of returned results and similarity score distribution, confirm they match the preset configuration.
- Test dynamic knowledge base calls, select variable reference in the "Knowledge Base Search" node, confirm the configured variable appears in the dropdown options.
- Trigger a rerank model test, check if the reranked order of retrieval results meets expectations, confirm the rerank function is enabled normally.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-14.