omni-ai-eval

Omni ships a first-class eval system (the AI Hub → Prompt sets and Eval runs). This skill drives it through the Omni CLI: define a reusable prompt set, start a judged eval run against a model or branch, and read per-prompt verdicts from Omni's built-in accuracy judge.

Prefer this native system over building your own harness. The judge scores each answer semantically against the full agent conversation — it does not require golden query JSON, and it evaluates the whole agentic workflow (topic selection, queries, results, and the final written answer), not just generated query structure.

Tip: Use omni-ai-optimizer to improve scores after finding failures, omni-model-builder to apply context changes on a branch before A/B testing, and omni-model-explorer to discover topics and fields when writing prompts.

# Verify the Omni CLI is installed — if not, ask the user to install it.
# See: https://github.com/exploreomni/cli#readme
command -v omni >/dev/null || echo "ERROR: Omni CLI is not installed."

# Verify the CLI has the eval commands. If this errors with "unknown command",
# the binary is stale — ask the user to update it (the ai-eval group is generated
# from the bundled API spec).
omni ai-eval --help >/dev/null 2>&1 || echo "ERROR: 'omni ai-eval' missing — update the Omni CLI."

# Show available profiles and select the right one — running against the wrong
# instance silently evaluates the wrong model.
omni config show
# If multiple profiles exist, ask the user which to use:
omni config use <profile-name>

# Confirm the active profile is authenticated and inspect your permissions:
omni whoami whoami

Auth: a profile authenticates with an API key or OAuth. If whoami (or any call) returns 401, hand off — ask the user to run ! omni config login <profile> (OAuth 2.1 browser flow; it blocks ~2 min on the browser). Don't run config login yourself in a headless/CI session (no browser → timeout); on a local interactive machine you may. See the omni-api-conventions rule for profile setup (omni config init --auth oauth) and discovering command and request-body shapes with --schema.

You also need the model ID of a shared model to evaluate. Evals require at least Querier access on that model, and at least one topic optimized for AI. See the Evals guide for concepts and prompt-set best practices.

All commands accept -o json (or --compact) to force structured output for parsing, and --profile <name> / --branch-id style global flags. Run omni ai-eval <command> --help for the full flag list. Run any command with --schema to print its args, flags, request-body schema (with a filled example) and response shape instead of guessing.

omni ai-eval prompt-sets-create --compact --body '{
  "model_id": "your-model-id",
  "name": "Orders regression",
  "slug": "orders-regression",
  "description": "Core revenue + orders coverage",
  "prompts": [
    { "prompt_text": "Show me revenue by month" },
    { "prompt_text": "What are the top 5 products by revenue?",
      "expectation": "The top product by revenue should be Aniseed Syrup." },
    { "prompt_text": "How many orders were placed last week?" }
  ]
}'

The response includes the created prompt_set.id (a UUID) — capture it for the run.

Find or update an existing set instead of recreating:

omni ai-eval prompt-sets-list --model-ids your-model-id --compact   # discover sets + ids
omni ai-eval prompt-sets-get <promptSetId> --compact                # full set with prompts

prompt-sets-update replaces the entire prompts list: omitted prompts are deleted, entries with no id are created, entries with a matching id are updated in place. To add one prompt, send the full desired list (existing prompts carry their id).

omni ai-eval prompt-sets-update <promptSetId> --compact --body '{
  "prompts": [
    { "id": "<existing-prompt-id>", "prompt_text": "Show me revenue by month" },
    { "prompt_text": "Revenue by month, last 12 months only" }
  ]
}'

# Against main:
omni ai-eval runs-create --compact --body '{
  "prompt_set_id": "<promptSetId>",
  "description": "Baseline on main"
}'

# Against a branch (measures a model-context change before promotion):
omni ai-eval runs-create --compact --body '{
  "prompt_set_id": "<promptSetId>",
  "description": "After adding ai_context to order_items",
  "run_config": { "branch_id": "<branchId>" }
}'

The response returns run.id and job_count (one agentic job per prompt). Omit run_config.branch_id to run against the live shared model.

Concurrency cap: at most 2 eval runs in flight at once. A 429 means the per-user active-run cap is reached — wait for an in-flight run to finish or cancel one. A 503 means eval is paused for the org. Check runs-list before launching.

runs-get returns results[], one row per prompt:

{
  "run": {
    "status": "COMPLETE",
    "branch_id": null,
    "results": [
      {
        "prompt": "What are the top 5 products by revenue?",
        "score": 1,
        "error_reason": null,
        "cost": 0.0021,
        "scoring_cost": 0.0004,
        "timing_ms": 4321,
        "agentic_job": { "state": "COMPLETE", "conversation_id": "conv-uuid", "id": "job-uuid" }
      }
    ]
  }
}

Overall accuracy = the pass rate (mean of score across results). Report it with the per-prompt breakdown, and for any failure, point to the conversation_id so the user can read why the judge failed it — that rationale is where the actionable signal lives.

Eval run "Baseline on main" — 9/12 passed (75.0%)
  ✗ "Revenue by quarter"        — judge: summed a row-limited result as a total
  ✗ "Top products this year"    — judge: date filter used calendar instead of fiscal year
  ✗ "Churn rate by segment"     — agentic job FAILED (error_reason)
  (open each conversation_id for the full rationale)

The core workflow for measuring whether a model change helps. Run the same prompt set twice — once on main, once on the branch — then compare.

1. Create or identify the branch with the change (use omni-model-builder to apply ai_context, fields, joins, ai_settings, etc. on a branch).
2. runs-create with no run_config → baseline run on main.
3. runs-create with run_config.branch_id → branch run.
4. Poll both to COMPLETE, then diff per-prompt verdicts.

A/B: main vs branch/new-context  (prompt set: orders-regression)
                    main      branch     Δ
Accuracy:           75.0%     91.7%     +16.7%
Prompt credits:     0.024     0.026     +0.002

Regressions (passed on main, failed on branch):
  - rev-by-quarter
Improvements (failed on main, passed on branch):
  - top-products-this-year
  - churn-rate-by-segment

Always check for regressions, not just net improvement. A higher overall pass rate can still hide a prompt that newly broke — an ai_context change that helps most prompts may conflict with one. Call out any prompt that passed on main but fails on the branch.

Notes for an auditable comparison:

• Both runs must use the same prompt set so they evaluate identical prompts (the AI Hub comparison view enforces this too).
• Record the full branch_id used and confirm the branch run's run.branch_id matches it — don't claim the branch was exercised without that.
• Expectations are snapshotted per run, so editing a prompt's expectation later won't change how earlier runs were scored.

omni ai-eval runs-list --prompt-set-id <promptSetId> --compact   # runs for a set, newest first
omni ai-eval runs-cancel <runId>                                 # cancel an in-flight run (also archives it)
omni ai-eval runs-archive <runId>                                # archive a finished run
omni ai-eval runs-unarchive <runId>                              # restore an archived run

omni ai-eval prompt-sets-archive <promptSetId>                   # archive a set (cancels its in-flight jobs)
omni ai-eval prompt-sets-unarchive <promptSetId>                 # restore an archived set

Archiving is a soft delete — sets and runs are preserved and can be restored. Cancelling a run marks it CANCELLED and archives it; use runs-unarchive to surface it in the default list again. Use --archived true on the *-list commands to see archived items.

• Right profile, right instance — the eval runs against whatever model lives on the active profile's instance. Confirm the profile before creating sets or runs.
• Judge is binary and fixed — a pass means the answer avoided the high-impact analysis errors, not that it's the single best possible answer. The judge model isn't configurable per run.
• Failed job ≠ failed judgment — a result with error_reason set means the agentic job itself failed (it never produced an answer to score); treat that separately from a judge score of 0.
• Snapshotting — to keep runs reproducible, capture the model state alongside results: omni models yaml-get <modelId> --compact and omni models validate <modelId>. Branch runs already pin the change to a branch.

• Evals (concepts + judge) · AI Eval API · Create AI job API · Optimizing models for AI

• omni-ai-optimizer — improve AI accuracy based on eval findings (ai_context, sample_queries, field metadata)
• omni-model-builder — apply context changes on a branch before A/B testing
• omni-model-explorer — discover topics and fields when writing prompts