LLM Eval Benchmark Harness

Writing Promptfoo Evals

You produce maintainable promptfoo eval suites: clear test cases, deterministic assertions where possible, model-graded only when needed.

See references/cheatsheet.md for the full assertion and provider reference. For deep questions about promptfoo features, consult https://www.promptfoo.dev/llms-full.txt

Inputs (infer from repo context if not provided)

• What is being evaluated (prompt, agent, endpoint, RAG pipeline)?
• What are the inputs and outputs (text, JSON, multi-turn chat, tool calls)?
• What does "good" look like (acceptance criteria, failure modes)?

If context is insufficient, scaffold with TODO markers and starter tests.

Workflow

1. Find or create the eval suite

Search for existing configs: promptfooconfig.yaml, promptfooconfig.yml, or any promptfoo/evals folder. Extend existing suites when possible.

For new suites, use this layout (unless the repo uses another convention):

evals/<suite-name>/
  promptfooconfig.yaml
  prompts/
  tests/

Always add # yaml-language-server: $schema=https://promptfoo.dev/config-schema.json at the top of config files.

2. Write prompts

• Put prompts in prompts/*.txt (plain) or prompts/*.json (chat format)
• Reference via file://prompts/main.txt
• Use {{variable}} for test inputs
• If the app builds prompts dynamically, use a JS/Python provider instead of duplicating logic

3. Choose providers

Pick the simplest option that matches the real system:

Keep provider count small: 1 for regression, 2 for comparison.

For JSON output, add response_format to the provider config:

config:
  temperature: 0
  response_format:
    type: json_object

4. Write tests

Use file-based tests so they scale: tests: file://tests/*.yaml

For larger suites, use dataset-backed tests:

tests: file://tests.csv
# or
tests: file://generate_tests.py:create_tests

Every test should have:

• description - short, specific
• vars - the inputs
• assert - validations (when automatable)

Cover: happy paths, edge cases, known regressions, safety/refusal checks, output format compliance.

5. Add assertions

Deterministic first (fast, reliable, free): equals, contains, icontains, regex, is-json, contains-json, starts-with, cost, latency, javascript, python

Model-graded sparingly (slow, costs money, non-deterministic): llm-rubric, factuality, answer-relevance, context-faithfulness

Assertions support optional weight (for scoring relative importance) and metric (named score in reports). threshold is assertion-specific: for graded assertions it is usually a minimum score (0-1), while for assertions like cost/latency it is a maximum allowed value.

For model-graded assertions, explicitly set the grader provider so grading is stable across runs:

defaultTest:
  options:
    provider: openai:gpt-5-mini

tests:
  - description: 'Model-graded quality check'
    assert:
      - type: llm-rubric
        value: 'Accurate and concise'
        # Optional per-assertion override:
        # provider: anthropic:messages:claude-sonnet-4-6

Hallucination / faithfulness pattern: When checking that output is grounded in source material, include the source in the rubric so the grader can compare. Use context-faithfulness when you have a context var, or inline the source in the llm-rubric value:

assert:
  - type: llm-rubric
    value: |
      The summary only states facts from this source article:
      "{{article}}"
      It does not add, infer, or fabricate any claims.

JSON output pattern:

assert:
  - type: is-json
    value: # optional JSON Schema
      type: object
      required: [name, score]
  - type: javascript
    value: 'JSON.parse(output).score >= 0.8'

Transform pattern (preprocess output before assertions): When models wrap JSON in markdown fences or add preamble text, use options.transform on the test to clean output before assertions run:

options:
  transform: "output.replace(/```json\\n?|```/g, '').trim()"

Use defaultTest for assertions shared across all tests (cost limits, format checks, etc.).

6. Validate and run

Before finishing, validate and provide run commands. Always use --no-cache during development to avoid stale results. Only run eval if credentials are available and safe to call.

npx promptfoo@latest validate config -c <config>
npx promptfoo@latest eval -c <config> -o output.json --no-cache --no-share

For CI/non-UI workflows, prefer the -o output.json command and inspect success, score, and error fields.

If working in the promptfoo repo itself, prefer the local build:

source ~/.nvm/nvm.sh && nvm use
npm run local -- validate config -c <config>
npm run local -- eval -c <config> -o output.json --no-cache --no-share

Add --env-file .env only when the eval needs local credentials and that file exists.

Do not run npm run local -- view unless explicitly asked.

Common mistakes

# ❌ WRONG — shell-style env vars don't work in YAML configs
apiKey: $OPENAI_API_KEY

# ✅ CORRECT — use Nunjucks syntax with quotes
apiKey: '{{env.OPENAI_API_KEY}}'

# ❌ WRONG — rubric references "the article" but grader can't see it
- type: llm-rubric
  value: 'Only contains info from the original article'

# ✅ CORRECT — inline the source so the grader can compare
- type: llm-rubric
  value: |
    Only states facts from: "{{article}}"

Output contract

When done, state:

• What the suite evaluates (1-3 bullets)
• Files created/modified (paths)
• How to run (copy-pastable commands)
• Required env vars
• TODOs left behind (only if unavoidable)