Eval datasets, curated by Bayes from your agents' real conversations

Bayes, Monte Carlo's AI teammate, now picks two eval datasets for every agent from the last 30 days of its scored conversations. Every example is a real question and the answer your agent gave:

• Golden: the top-scoring answers. Use it to catch regressions, because a new version has to match these.
• Improvement: the lowest-scoring answers. Use it to measure fixes, because a new version has to beat these.

Examples are spread across your agent's intent clusters, so they cover the kinds of questions your users ask, not just the most common one. Improvement also covers its issue clusters.

Found a great answer while debugging, or a failure you want every new version to beat? Select Pin on any single-turn conversation to add it to Golden or Improvement yourself, whatever its score.

On your agents 'Datasets' card, open either set to preview each example exactly as it will be exported, uncheck the ones you don't want, and export in the format your eval tool already uses: a table for Snowflake Cortex evaluations, benchmark questions for Databricks AI/BI Genie, an eval set for Google ADK, or JSONL for any other eval tool.

Each example comes with tips for turning it into a strong test case, linked to your platform's best practices.

Learn more here: Eval datasets