Recommending a fix when a data source's text is problematic

When a GenAI application's data source contains problematic text (e.g., toxic, biased, PII-laden, or otherwise unsafe content), the recommended mitigation is to filter, redact, or replace that content before it enters the retrieval or fine-tuning pipeline rather than relying solely on downstream guardrails. Governance in Databricks favors addressing data quality at the source using Unity Catalog controls, curated/cleaned tables, and permissioned access rather than attempting to patch issues only at inference time.

1 · Learn the must-know

• 1Filtering or cleansing problematic text should happen upstream in the data pipeline (e.g., during ETL into a curated Delta table) rather than only being caught by a runtime guardrail model.
• 2Unity Catalog can be used to restrict access to raw or sensitive source tables, exposing only a governed, cleaned, or masked version to the RAG or fine-tuning pipeline.
• 3Techniques such as column masking, row filtering, or PII redaction functions in Unity Catalog can be applied to remediate sensitive or problematic content before it is indexed or embedded.
• 4Relying exclusively on a downstream safety/guardrail layer (like a moderation model on outputs) is not a substitute for fixing or removing problematic content at the data source, since it treats symptoms rather than the root cause.
• 5A recommended alternative when a source is inherently unreliable or contains too much problematic content is to replace it with a curated, higher-quality, well-governed data source rather than attempting extensive inline filtering.
• 6Data lineage and audit logging in Unity Catalog help track which source produced problematic content, supporting root-cause remediation and compliance reporting.

2 · Check your understanding