Truthound local CSV and Parquet use cases

“Check this CSV for missing values and duplicate rows.” Select check_file after directory approval. A successful tool execution can report a failed data-quality check; do not turn that finding into a transport error or a quality pass.

“Profile this Parquet file before RAG ingestion.” Select profile_file for inferred types and aggregate column statistics. Observed types are not validation against an expected schema. Use the truncation fields to understand bounded output.

“Find likely PII in a local dataset.” Select scan_pii for sampled findings before sharing or ingestion. Counts are estimates. A scan with no findings does not prove PII is absent, and does not certify compliance.

“Validate my evaluation dataset without uploading it.” Ask which checks are intended; null/duplicate inspection and PII screening of a local tabular file are supported. Label correctness, model performance, document relevance and RAG answer faithfulness need other evaluators.

For Pandera DataFrameSchema, Great Expectations suites or Soda SQL rules, follow the requested framework and its actual contract. The Truthound Local plugin does not implement their schemas or execute their configurations. Database migrations, ETL transformations, cloud DLP, charts and dataset editing are outside this surface.