본문으로 건너뛰기

Evaluation Guide

Understanding and interpreting FluxLoop test evaluation results.

Overview​

FluxLoop evaluates agent test results using:

  • Automated criteria scoring
  • LLM-based evaluation
  • Performance metrics
  • Behavioral analysis

Evaluation Process​

1. Run Tests​

fluxloop test --scenario production --no-skip-upload

2. Automatic Evaluation​

Results are evaluated against criteria:

# Evaluation criteria
criteria:
  - accuracy: "Factual correctness"
  - helpfulness: "User helpfulness"
  - safety: "No harmful content"

3. View Results​

# View locally
fluxloop results show

# View on Web Platform
# https://results.fluxloop.ai/run/run_123

Evaluation Criteria​

LLM Judge​

LLM evaluates subjective qualities:

criteria:
  - id: accuracy
    type: "llm_judge"
    weight: 0.4
    rubric: |
      1.0: Completely accurate
      0.7: Mostly accurate, minor errors
      0.4: Significant inaccuracies
      0.0: Incorrect information

Metrics​

Automated performance measurements:

criteria:
  - id: response_time
    type: "metric"
    weight: 0.2
    thresholds:
      excellent: 500   # < 500ms
      good: 1000       # < 1s
      acceptable: 2000 # < 2s

Classifiers​

Binary pass/fail checks:

criteria:
  - id: safety
    type: "classifier"
    weight: 0.4
    required: true  # Must pass

Scoring​

Individual Scores​

Each criterion scored 0.0 - 1.0:

{
  "input_id": "input_001",
  "scores": {
    "accuracy": 0.9,
    "helpfulness": 0.8,
    "safety": 1.0
  }
}

Overall Score​

Weighted average:

overall_score = (accuracy * 0.4) + (helpfulness * 0.4) + (safety * 0.2)
              = (0.9 * 0.4) + (0.8 * 0.4) + (1.0 * 0.2)
              = 0.88

Pass/Fail​

Based on threshold:

evaluation:
  pass_threshold: 0.75

# Score 0.88 >= 0.75 → PASS ✅

Interpreting Results​

High Score (0.9+)​

✅ Excellent performance

  • Agent meets or exceeds expectations
  • Ready for production

Good Score (0.75-0.9)​

✅ Good performance

  • Acceptable for production
  • Minor improvements possible

Marginal Score (0.6-0.75)​

⚠️ Needs improvement

  • Some issues present
  • Review failures
  • Iterate on prompts

Poor Score (< 0.6)​

❌ Significant issues

  • Not ready for production
  • Major revisions needed

Common Patterns​

High Accuracy, Low Helpfulness​

accuracy: 0.9
helpfulness: 0.5

Agent provides correct but unhelpful responses.

Fix: Improve response formatting, add examples.

High Helpfulness, Low Accuracy​

accuracy: 0.6
helpfulness: 0.9

Agent is friendly but provides wrong information.

Fix: Improve knowledge base, add fact-checking.

Inconsistent Scores​

input_001: 0.9
input_002: 0.4
input_003: 0.9

Agent handles some cases well, others poorly.

Fix: Identify failure patterns, add specific training.

Best Practices​

1. Define Clear Criteria​

✅ Good:
  - id: accuracy
    description: "Provides factually correct information"

❌ Vague:
  - id: quality
    description: "Good response"

2. Use Multiple Criteria​

Don't rely on single metric:

criteria:
  - accuracy
  - helpfulness
  - safety
  - response_time

3. Regular Evaluation​

Evaluate frequently:

# Daily in CI
fluxloop test --scenario regression --no-skip-upload

Compare results over time:

fluxloop results compare --baseline last-week

Next Steps​

  • Define evaluation criteria for your agent
  • Run tests and review scores
  • Iterate based on results
  • Set up continuous evaluation