create-eval

Creating HolmesGPT Eval Tests

This skill provides the complete workflow for creating LLM evaluation tests in the HolmesGPT project. Eval tests validate that Holmes can correctly answer questions by querying real infrastructure and services.

Test Structure

Each eval lives in its own directory under tests/llm/fixtures/test_ask_holmes/:

tests/llm/fixtures/test_ask_holmes/<NNN>_<descriptive_name>/
├── test_case.yaml          # Required: test definition
├── toolsets.yaml            # Optional: enable specific toolsets
├── manifest.yaml            # Optional: Kubernetes manifests
├── generate_*.py            # Optional: data generation scripts
└── other supporting files

Naming convention: <3-digit-number>_<snake_case_description> (e.g., 212_large_configmap_needle).

Creation Workflow

Step 1: Choose Test Number and Namespace

Check existing tests to find the next available number:

ls tests/llm/fixtures/test_ask_holmes/ | sort -n | tail -5

The namespace must be app-<testid> (e.g., app-212). All pod and resource names must be unique across all tests.

Step 2: Validate Tags

Only use tags that exist in pyproject.toml markers section. Using invalid tags causes test collection failures. Read pyproject.toml and check the [tool.pytest.ini_options] markers list before assigning tags. Ask the user before adding any new tag.

Step 3: Write test_case.yaml

Core fields:

user_prompt: "Specific question for Holmes to answer"

expected_output:
  - "Criterion 1: Must report exact value X"
  - "Criterion 2: Must include specific identifier Y"

tags:
  - kubernetes
  - question-answer
  - hard

before_test: |
  set -e
  # Setup infrastructure...

after_test: |
  kubectl delete namespace app-NNN --ignore-not-found

For the complete field reference and all available options, consult references/test-case-format.md.

Step 4: Write toolsets.yaml (if needed)

When the test requires specific toolsets (Prometheus, Grafana, Elasticsearch, etc.):

toolsets:
  kubernetes/core:
    enabled: true
  prometheus/metrics:
    enabled: true
    config:
      prometheus_url: http://localhost:10033

When a toolsets.yaml exists, only explicitly enabled toolsets are available to the LLM. All others are disabled.

Step 5: Write Setup Scripts

The before_test script runs from the test's directory via /bin/bash. Key rules:

• Always start with set -e to fail on any error
• Use kubectl create namespace app-NNN --dry-run=client -o yaml | kubectl apply -f - for idempotent namespace creation
• Use exit 1 when verification fails to fail the test early
• Clean up temp files at the end of before_test

Verification focus: verify the needle, not the haystack. The only verification that matters is that Holmes can discover the answer. Run the same kind of query Holmes would run and check that the expected value (the "smoking gun") is present. Do NOT exhaustively verify every piece of infrastructure — if the needle is queryable, the environment is working. Keep setup scripts short and readable.

# GOOD - verify the needle is discoverable (one targeted check)
kubectl get configmap platform-config -n app-212 \
  -o jsonpath='{.data.platform-config\.yaml}' | grep -q '7k3m9x'

# BAD - verifying everything (pod health, service endpoints, API responses, readiness...)
# This bloats the script without adding value

For retry loop patterns and other infrastructure details, consult references/infrastructure-patterns.md.

Step 6: Design Anti-Hallucination Measures

Every eval must be designed so the LLM cannot pass by guessing. This is the most critical aspect of eval design.

Key principles:

• Embed unique random identifiers that cannot be guessed (e.g., 7k3m9x)
• Test for specific values discoverable only by querying
• Use neutral resource names that don't hint at the problem
• Write prompts that test discovery ability, not domain knowledge

For detailed anti-hallucination patterns and examples, consult references/anti-hallucination.md.

Mandatory Testing Workflow

Always run evals before submitting when possible. Follow this sequence:

Phase 1: Verify Collection

poetry run pytest -k "test_name" --collect-only -q --no-cov

Confirm the test appears in the output. If not, check for tag or YAML errors.

Phase 2: Run Setup Only

poetry run pytest -k "test_name" --only-setup --no-cov

Verify setup completes without errors. Check that infrastructure is ready.

Phase 3: Run Full Test

poetry run pytest -k "test_name" --no-cov --skip-setup

Use --skip-setup to reuse the infrastructure from Phase 2. Verify the test passes.

Phase 4: Verify Cleanup

kubectl get namespace app-NNN

Should return NotFound after the test completes (unless --skip-cleanup was used).

Key Rules

1. Namespace isolation: Every test uses app-<testid> namespace
2. Unique resource names: Never reuse pod/service names across tests
3. No :latest tags: Always use specific container image versions
4. Secrets for scripts: Use Kubernetes Secrets for scripts, not ConfigMaps or inline
5. No hints in names: Avoid broken-pod, crashloop-app — use neutral names
6. Sign commits: Always use git commit -s for DCO compliance

Quick Reference: Common Patterns

Additional Resources

Reference Files

For detailed documentation, consult:

• references/test-case-format.md — Complete test_case.yaml field reference with all options
• references/anti-hallucination.md — Anti-cheat testing patterns and prompt design
• references/infrastructure-patterns.md — Setup scripts, retry loops, port forwards, shared infra
• references/running-evals.md — CLI flags, environment variables, model comparison, debugging