May 22, 2026: AI_EXTRACT extraction scores (General availability)

Extraction scores for the Snowflake Cortex AI_EXTRACT function are now generally available. Key capabilities include:

  • Confidence scores for extracted values: Pass scores => TRUE in an AI_EXTRACT call to get back a scoring object alongside the standard response object, with a score between 0 and 1 indicating the model’s certainty about each extracted value.
  • Support for entities, lists, and tables: Per-field scores are returned for entity extraction, and aggregate scores are returned for list and table extraction.
  • Support for fine-tuned arctic-extract models: Extraction scores work with fine-tuned arctic-extract models in addition to the default model.
  • Human-in-the-loop workflows: Use scores to set thresholds and flag low-scoring extractions for human review, and to build fallback mechanisms for unreliable extractions.
  • No additional cost: Requesting scores does not incur additional cost beyond the standard AI_EXTRACT call.

For more information, see Extraction scores.