Agents over structured data
Cortex Agents that use Cortex Analyst to turn questions into SQL over your semantic views. Metric definitions and joins decide whether answers are right.
SERVICES / DATA PLATFORMS
Evaluate Cortex Agents on the answers your business relies on.
Independent testing and evaluation for AI agents built with Snowflake Cortex Agents: tool selection, text-to-SQL accuracy, search grounding and answer correctness.
Common Snowflake agent patterns
Agents over documents
Agents that use Cortex Search to retrieve from unstructured content. Retrieval quality shapes every answer and citation.
Agents that combine both
Agents that plan across Analyst, Search and custom tools, then answer in Snowflake CoWork (formerly Snowflake Intelligence) or your own application.
Quality issues we look for
Wrong tool for the question
The agent searches documents when the answer needs data, or queries data when the answer is in a policy document.
Plausible but wrong numbers
SQL with the wrong metric, filter, time range or join, often caused by gaps in the semantic view, returned with full confidence.
Unsupported answers and citations
Answers that go beyond retrieved content, or citations that don't support the claim they are attached to.
What we assess
Answer correctness
Agent answers compared with reviewed expected answers for representative business questions, including ambiguous and unanswerable ones.
Tool selection and execution
Whether the agent chooses the right tools, in a sensible plan, and whether each tool call returns what the question needed.
Semantic model coverage
Which questions fail because of the semantic view rather than the agent, so fixes go to the right place.
Working with native Snowflake evaluation
Cortex Agent evaluations
Where available to your account, we use Snowflake's agent evaluations, including answer correctness and tool selection and execution accuracy, with datasets reviewed by your team.
Traces as evidence
AI Observability traces show the plan, tool calls and results behind each answer, so findings can be reproduced and explained.
Calibrated against your analysts
Automated scores are compared with your analysts' judgment before they inform a release, and agent configurations are recorded with every run.
Related services
Evaluation Assessment
A baseline for one Cortex agent or workflow, with prioritized findings and a release recommendation.
Explore serviceHallucination Testing
Measure how often your Cortex agent invents facts, policies or actions, and which fixes reduce it.
Explore serviceManaged Evaluations
Recurring tests as your Cortex agent, prompts, models and data change, with a monthly quality report.
Explore servicePlan the engagement
Your inputs
An account and role with agreed access, agent and semantic view versions, representative questions with expected answers and a data owner who can review results.
Your deliverables
An agreed benchmark, prioritized findings with reproducible cases, a reusable evaluation dataset and recommendations for the next iteration.
Support after handover
We can retest fixes, extend coverage to new Cortex workflows or move to recurring Managed Evaluations as your agent changes. Platform usage and licenses stay in your account.
Make your next decision with evidence.
Start with one workflow, clear success criteria, and an evaluation plan your team can use.