06 Evaluation

Track A Step 6: Evaluate the Agent

Verify your agent produces relevant, grounded, and safe responses before deploying it. The template includes evaluation infrastructure — this step teaches you to run it, interpret results, and customize the dataset.

When to Use

• Your agent is running locally with tools, auth, and (optionally) memory configured from Steps 1-5.
• You want to validate response quality before deploying to Databricks Apps.
• You need to establish a baseline for ongoing monitoring after deployment.

Relationship to SDLC Evaluation

This step runs the template's built-in evaluation as a smoke test. The full SDLC evaluation pipeline (S2–S4) adds custom benchmark datasets, domain-specific scorers with threshold gates, and human feedback labeling. Think of A6 as "does my agent work?" and S4 as "does my agent meet production quality bars?"

Step 6a: Run the Built-In Evaluation

Smoke evaluation uses the configured model route from docs/agent_tool_plan.yaml.runtime_config.llm. The evaluation runner must not substitute a hardcoded model endpoint. If the route is the default Databricks provider, this resolves to llm_endpoint in config.yml.

The agent template includes evaluation code in agent_server/evaluate_agent.py (or tests/evaluate_agent.py depending on your template version). Run it:

uv run agent-evaluate

What this does:

1. Loads the evaluation dataset (typically tests/eval_dataset.json)
2. Sends each test case to your agent via predict()
3. Runs built-in judges against the responses
4. Outputs a summary report with pass/fail and scores

Expected output:

Running evaluation...
  ✓ relevance: 4/5 passed
  ✓ groundedness: 5/5 passed
  ✓ safety: 5/5 passed
Evaluation complete. Results saved to MLflow experiment.

If the command fails with auth errors, see the auth-in-evaluation section below.

Step 6b: Understand the Evaluation Dataset

The template ships a sample dataset in tests/eval_dataset.json:

[
  {
    "request": {
      "input": [
        {"role": "user", "content": "What is MLflow?"}
      ]
    },
    "expected_response": "MLflow is an open source platform for managing the ML lifecycle.",
    "expected_retrieved_context": [
      {"doc_uri": "https://mlflow.org/docs/latest/index.html"}
    ]
  }
]

The expected_retrieved_context field connects to set_retriever_schema() from Step 3 — the doc_uri column is how judges match retrieved vs. expected documents.

Step 6c: Understand the Built-In Judges

Agent Evaluation runs multiple judges automatically:

Judges are LLM-based — they use a Databricks model serving endpoint to evaluate quality. The evaluation harness handles this automatically.

Step 6d: Interpret Results

After running evaluation, results appear in two places:

1. Terminal output

Summary pass/fail counts and aggregate scores.

2. MLflow Experiment UI

Navigate to your experiment → Evaluation tab for:

• Per-request scores and judge rationale
• Trace visualization for each test case
• Aggregate metrics across the dataset

Look for:

• Low relevance scores → Agent may need better instructions or tools
• Low groundedness → Agent may be hallucinating beyond retrieved context
• Safety failures → Agent needs guardrails or instruction updates

Step 6e: Customize the Evaluation Dataset

Replace the sample dataset with test cases for your agent's domain:

[
  {
    "request": {
      "input": [
        {"role": "user", "content": "What were Q3 2025 total sales?"}
      ]
    },
    "expected_response": "Q3 2025 total sales were $42.3M.",
    "expected_retrieved_context": [
      {"doc_uri": "catalog.schema.sales_summary"}
    ]
  },
  {
    "request": {
      "input": [
        {"role": "user", "content": "Who is the VP of Engineering?"}
      ]
    },
    "expected_response": null
  }
]

Guidelines for good eval datasets:

• Include 10-50 test cases covering your agent's core capabilities
• Mix questions that require tools with general knowledge questions
• Include edge cases (ambiguous queries, out-of-scope questions)
• Set expected_response to null when you only want safety/groundedness checks (no relevance baseline)
• Use real doc_uri values from your Vector Search indexes

Step 6f: Run Evaluation Programmatically

For more control, call mlflow.evaluate() directly:

import mlflow
import json

eval_dataset = json.load(open("tests/eval_dataset.json"))

with mlflow.start_run():
    results = mlflow.evaluate(
        model="agent.py",
        data=eval_dataset,
        model_type="databricks-agent",
    )

    print(f"Metrics: {results.metrics}")
    print(f"Results table:\n{results.tables['eval_results']}")

This gives you access to:

• results.metrics — aggregate scores (relevance, groundedness, safety)
• results.tables["eval_results"] — per-row results with judge rationale

Authentication in Evaluation Context

Evaluation runs use different auth than Model Serving or Databricks Apps. Key differences:

If your agent uses OBO (IS_IN_DB_MODEL_SERVING_ENV), evaluation will fail because OBO environment variables are absent. Your agent's environment detection from Step 4 handles this — it falls back to default WorkspaceClient() auth outside Model Serving.

If evaluation still fails with permission errors:

1. Verify your agent declares resources in SystemAuthPolicy (see Step 4, references/model-serving-auth-passthrough.md)
2. Ensure the evaluation user has access to all resources the agent needs
3. Check that Genie Spaces and Vector Search indexes are granted to the system SP

DO / DON'T

DO — Run evaluation before every deployment

uv run agent-evaluate
# Review results, then deploy

DON'T — Deploy without evaluating

Changes to instructions, tools, or models can cause regressions that are invisible without evaluation.

DO — Keep the eval dataset in version control

tests/
├── eval_dataset.json          # Your test cases
└── eval_dataset_sample.json   # Template's original (for reference)

DON'T — Use only the template's sample dataset

The sample covers generic cases. Replace it with your domain-specific test cases for meaningful quality assurance.

DO — Treat low scores as blockers

# If relevance < 80% or safety has failures:
# 1. Review failing test cases in MLflow UI
# 2. Fix agent instructions or tools
# 3. Re-run evaluation

Common Issues

Validation Gate

All must pass before proceeding:

• uv run agent-evaluate runs to completion without errors
• Evaluation results visible in MLflow Experiment UI
• Relevance scores meet your threshold (recommended: ≥80%)
• Zero safety failures
• Evaluation dataset customized with domain-specific test cases
• Auth works in evaluation context (no permission errors)

Next Step

After passing this gate, load and execute tracks/A-custom-agent-apps/07-deploy-and-query/SKILL.md (A7: Deploy and Query) for the first manual deployment.

Going Further: Full SDLC Evaluation

The smoke test above validates basic functionality. After completing Track A (A7), you will enter the SDLC pipeline to build the comprehensive evaluation with custom scorers and threshold gates. When you reach that point, load these skills in order:

1. Load and execute sdlc/02-evaluation-datasets/SKILL.md (S2) — build your benchmark dataset (50-500+ rows)
2. Load and execute sdlc/03-scorers-and-judges/SKILL.md (S3) — create domain-specific scorers with pass/fail thresholds
3. Load and execute sdlc/04-evaluation-runs/SKILL.md (S4) — run mlflow.genai.evaluate() with your predict_fn from A5

S4 uses the same predict_fn you produced in A5 — no additional wrapping needed. The SDLC evaluation replaces this smoke test as the production quality gate.

Notes to Carry Forward

Related Skills

References

• Author an agent — Evaluate the agent
• Agent Evaluation
• Evaluation sets
• MLflow evaluate()
• Agent Evaluation judges

Version History