Skip to content

README.md

Latest commit

 

History

History
395 lines (278 loc) · 10.3 KB

README.md

File metadata and controls

395 lines (278 loc) · 10.3 KB
Raw

AI Evaluator CLI — Python

PyPI Python

Evaluate your LLM agents from the terminal. No browser. No dashboard.

pip install aievaluator

🧭 Tutorial — From Zero to CI/CD

Every step builds on the previous one. Start wherever makes sense for you.

Level 0 — Try it without installing anything

curl -s -X POST https://api.aievaluator.dev/api/v1/playground/evaluate \
  -H "Content-Type: application/json" \
  -d '{"queries":["What is 2+2?"],"metrics":["faithfulness"]}' | jq .

5 free per day. No key. No install. Good enough to decide if it's useful.

Level 1 — Install and evaluate a single prompt

pip install aievaluator

# Ask a question, tell it what you expect
aievaluator quick "What is the capital of France?" --expected "Paris"

You'll see a table with the score. The --expected is optional — without it, the judge evaluates the response on its own merits.

⚠️  Playground mode — 4/5 remaining

  AI Evaluator — Results
  Overall Score:  95.0%  ✅ above threshold (0%)
  Total rows:     1
  Failed:         0

┌────┬────────────────────────────────────┬──────────┬──────┐
│  # │ Query                              │ Score    │ Pass │
├────┼────────────────────────────────────┼──────────┼──────┤
│  1 │ What is the capital of France?     │  95%     │ ✅   │
└────┴────────────────────────────────────┴──────────┴──────┘

Level 2 — Sign up and scaffold a project

Playground is great for trying, but you'll want more than 5 evals/day.

# Get your API key at https://aievaluator.dev/settings
aievaluator login

# Check your account
aievaluator whoami

Now scaffold your project:

aievaluator init

This creates:

  • aievaluator.config.json — project-local config
  • evals/smoke-test.json — sample dataset with 3 queries
  • Updates .gitignore

Open evals/smoke-test.json and replace the sample queries with your own:

[
  {"input": "What are your business hours?", "expected_output": "Mon-Fri 9am-6pm"},
  {"input": "How do I cancel my order?", "expected_output": "Go to My Orders → Cancel"},
  {"input": "Do you ship internationally?", "expected_output": "Yes, via DHL Express"}
]

Test it against the built-in agent:

aievaluator quick --dataset ./evals/smoke-test.json

Level 3 — Evaluate your own agent

Point the CLI at your agent's endpoint:

aievaluator eval \
  --agent https://chatbot-staging.acme.com/api/chat \
  --dataset ./evals/smoke-test.json \
  --metrics faithfulness,g_eval

The CLI calls your agent with each query, then an LLM judge scores the responses.

Level 4 — Add quality gates

Not all metrics are equally important. Set different thresholds per metric:

aievaluator eval \
  --agent https://chatbot-staging.acme.com/api/chat \
  --dataset ./evals/smoke-test.json \
  --thresholds faithfulness:0.90,g_eval:0.75
  • faithfulness must be ≥ 90% (hallucination = instant fail)
  • g_eval must be ≥ 75% (general quality)

If any metric fails to meet its threshold, that row is marked ❌.

Or set one bar for everything:

aievaluator eval \
  --agent https://chatbot-staging.acme.com/api/chat \
  --dataset ./evals/smoke-test.json \
  --min-score 0.80

This works on quick too:

aievaluator quick "test prompt" --min-score 0.80
# Exit code 1 if any metric drops below 0.80

Level 5 — Create your own evaluation criteria

Sometimes the built-in metrics aren't enough. Define a custom evaluator inline:

aievaluator eval \
  --agent https://chatbot-staging.acme.com/api/chat \
  --dataset ./evals/smoke-test.json \
  --metrics politeness,g_eval \
  --custom '{"name":"politeness","prompt":"Is the response polite and professional? Answer YES or NO and explain.","threshold":0.85}'

The custom evaluator politeness is defined in the request, referenced in --metrics by name, and evaluated alongside g_eval. No dashboard needed.

Custom evaluator with per-metric threshold override:

aievaluator eval \
  --agent $URL --dataset ./tests.json \
  --metrics politeness,g_eval \
  --custom '{"name":"politeness","prompt":"Is the tone friendly?","threshold":0.7}' \
  --thresholds politeness:0.90,g_eval:0.80

The --thresholds flag overrides whatever was set in --custom. The engine uses the per-evaluation value.

Level 6 — CI/CD pipeline

Add this to your GitHub Actions, GitLab CI, or Jenkins:

aievaluator eval \
  --agent $STAGING_AGENT \
  --dataset ./evals/regression.json \
  --thresholds faithfulness:0.90,g_eval:0.75 \
  --min-score 0.80 \
  --ci \
  --format junit > report.xml
Flag What it does
--ci No colors, no prompts — clean output for logs
--format junit JUnit XML that CI systems understand natively
--min-score 0.80 Overall score must be ≥ 80%
--thresholds Per-metric quality bars

Exit code 1 = pipeline fails = deploy blocked.

If your agent requires authentication:

aievaluator eval \
  --agent $STAGING_AGENT \
  --agent-auth-type bearer \
  --agent-auth-token $AGENT_BEARER_TOKEN \
  --dataset ./evals/regression.json \
  --min-score 0.80 \
  --ci \
  --format junit > report.xml

Supported auth types: none (default), api_key, bearer. For api_key, the default header is X-API-Key. For bearer, the default header is Authorization with Bearer <token> format. Override with --agent-auth-header.

Environment variables for CI:

export AIEVALUATOR_API_KEY="sk-..."       # No hardcoded keys in YAML
export AIEVALUATOR_ENGINE_URL="https://api.aievaluator.dev"

📋 Complete Command Reference

aievaluator login

aievaluator login                        # Interactive prompt
aievaluator login --api-key sk-xxx       # Non-interactive (CI)
aievaluator login --engine-url https://custom.engine.com

aievaluator whoami

aievaluator whoami
# Tenant:  acme-corp
# Tier:    pro
# Evals:   42/5000 this cycle
# Tokens:  ↓124,800 · ↑89,200 this cycle

aievaluator quick

# Single query
aievaluator quick "What is 2+2?" --expected "4"

# Per-metric thresholds
aievaluator quick "test" --metrics faithfulness:0.90,g_eval:0.75

# General threshold
aievaluator quick "test" --min-score 0.80

# From dataset (JSON or JSONL)
aievaluator quick --dataset ./tests.json
aievaluator quick --dataset ./tests.jsonl

# Custom judge model
aievaluator quick "test" --judge deepseek

# With agent authentication
aievaluator quick --dataset ./tests.json \
  --agent-auth-type bearer --agent-auth-token "sk-abc123"

aievaluator eval

# Basic
aievaluator eval --agent $URL --dataset ./tests.json

# With agent authentication (Bearer token)
aievaluator eval --agent $URL --dataset ./tests.json \
  --agent-auth-type bearer --agent-auth-token "sk-abc123"

# With agent authentication (API key)
aievaluator eval --agent $URL --dataset ./tests.json \
  --agent-auth-type api_key --agent-auth-header "X-API-Key" --agent-auth-token "my-secret"

# With quality gates
aievaluator eval --agent $URL --dataset ./tests.json \
  --thresholds faithfulness:0.90,g_eval:0.75 --min-score 0.80

# Inline rows
aievaluator eval --agent $URL \
  --rows '[{"input":"Hi","expected_output":"Hello"}]'

# Custom evaluator inline
aievaluator eval --agent $URL --dataset ./tests.json \
  --metrics my-eval --custom '{"name":"my-eval","prompt":"...","threshold":0.8}'

# CI mode
aievaluator eval --agent $URL --dataset ./tests.json --ci --format junit

# Different agent format
aievaluator eval --agent $URL --dataset ./tests.json --agent-format claude

# Agent authentication (API key or Bearer token)
aievaluator eval --agent $URL --dataset ./tests.json \
  --agent-auth-type bearer --agent-auth-token "sk-abc123"

aievaluator eval --agent $URL --dataset ./tests.json \
  --agent-auth-type api_key --agent-auth-header "X-API-Key" --agent-auth-token "my-key"

aievaluator config

aievaluator config show
aievaluator config set default-metrics "faithfulness,g_eval"
aievaluator config set default-min-score 0.80
aievaluator config unset default-min-score

aievaluator init

aievaluator init
# Creates aievaluator.config.json + evals/smoke-test.json + updates .gitignore

aievaluator generate-ci

Generates a CI/CD workflow file for GitHub Actions or GitLab CI.

aievaluator generate-ci --platform github

Options:

Flag Default Description
--platform github|gitlab github CI/CD platform
--dataset ./evals/regression.json Dataset path
--output stdout Save to file 
# Print GitHub Actions workflow
aievaluator generate-ci --platform github

# Save GitLab CI workflow to file
aievaluator generate-ci --platform gitlab --output .gitlab-ci.yml

📊 Output Formats

Table (default)

Human-readable table with scores, pass/fail icons, and token counts.

JSON (--format json)

aievaluator eval ... --format json | jq '.overall_score'

Clean JSON on stdout. All logs/warnings go to stderr.

JUnit XML (--format junit)

aievaluator eval ... --format junit > report.xml

Native CI integration. <testcase> per query, <failure> for queries below threshold.

🌐 Evaluating local agents

Use --tunnel to evaluate agents running on localhost or private networks:

aievaluator eval --agent http://localhost:8047/chat --tunnel --dataset ./tests.json
aievaluator quick "Hello" --agent http://localhost:8047/chat --tunnel

The CLI auto-detects local URLs and creates a public tunnel via cloudflared (free, no signup), ngrok, bore, or localtunnel. The tunnel is closed automatically when the evaluation finishes.

Install cloudflared: brew install cloudflared

Requirements

  • Python 3.10+