Skip to main content

Documentation Index

Fetch the complete documentation index at: https://www.adaline.ai/docs/llms.txt

Use this file to discover all available pages before exploring further.

LLM-as-a-Judge is the most versatile evaluator in Adaline. It uses an LLM to assess your prompt outputs against a custom rubric, excelling at qualitative assessment where nuanced judgment matters more than simple metrics.

​
Set up LLM-as-a-Judge

1

Select the evaluator

Add the LLM-as-a-Judge evaluator from the evaluator menu.Adding the LLM-as-a-Judge evaluator
2

Link a dataset

Give a name to the evaluator and link a dataset containing your test cases.Linking a dataset to LLM-as-a-Judge
3

Define your rubric

Write the rubric that defines your evaluation criteria. A well-crafted rubric is the key to high-quality evaluations — it should be specific, actionable, and aligned with your success metrics.Defining a rubric for LLM-as-a-Judge
4

Run the evaluation

Click Evaluate to execute the evaluation and see the results.LLM-as-a-Judge evaluation results

​
Writing effective rubrics

Your rubric directly determines the quality of the evaluation. Follow these guidelines:
  • Be specific — Define clear criteria for each score level. Avoid vague terms like “good” or “bad” without context.
  • Be actionable — Include concrete examples of what constitutes each rating.
  • Align with goals — Match your rubric to your actual production success metrics.
  • Use a consistent scale — Define a numerical scale (e.g., 1–4 or 1–5) with clear definitions for each level.

​
Example rubrics

​
Customer support response quality

Evaluate this customer support response using the following criteria:

Scoring Scale (1-4):
4 - Excellent: Completely resolves the issue, professional tone, anticipates follow-up needs
3 - Good: Addresses the main concern clearly and professionally
2 - Fair: Partially helpful but missing key information or context
1 - Poor: Fails to address the issue or uses inappropriate tone

Evaluation Factors:
- Problem resolution completeness
- Professional communication standards
- Information accuracy
- User experience quality

Provide a score and brief justification for your assessment.

​
Content marketing effectiveness

Rate this content piece on effectiveness for our target audience (1-5):

5 - Outstanding: Highly engaging, actionable insights, clear value proposition
4 - Strong: Good engagement with solid practical value
3 - Adequate: Informative but limited engagement or actionability
2 - Weak: Basic information with minimal practical value
1 - Poor: Lacks clarity, value, or relevance to target audience

Consider these dimensions:
- Audience alignment and relevance
- Practical value and actionability
- Engagement potential
- Brand positioning effectiveness

​
Technical documentation quality

Assess this feature documentation quality (1-4):

4 - Comprehensive: Clear explanation, complete coverage, excellent user guidance
3 - Good: Well-explained with adequate detail and guidance
2 - Acceptable: Basic explanation but missing important details or clarity
1 - Inadequate: Confusing, incomplete, or lacks necessary user guidance

Evaluation Areas:
- Technical accuracy and completeness
- User comprehension and clarity
- Implementation guidance quality
- Overall user experience

​
Brand voice consistency

Evaluate brand voice alignment (1-3 scale):

3 - Excellent Alignment: Perfect adherence to brand guidelines, authentic voice
2 - Good Alignment: Generally consistent with minor deviations
1 - Poor Alignment: Inconsistent with established brand voice

Assessment Criteria:
- Tone consistency with brand guidelines
- Language and terminology alignment
- Audience appropriateness
- Brand personality expression

​
When to use

LLM-as-a-Judge is best suited for:
  • Response quality assessment — Evaluating helpfulness, accuracy, and completeness.
  • Tone and voice validation — Ensuring consistent brand voice and appropriate communication style.
  • Factual accuracy checking — Verifying that responses contain correct information.
  • User satisfaction prediction — Assessing whether responses would meet user expectations.
  • Multi-turn conversation quality — Evaluating context retention and coherence across chat turns.
For technical validation (format checking, business logic, pattern matching), consider JavaScript or Text Matcher evaluators instead.

​
Next steps

JavaScript Evaluator

Write custom code to validate structured outputs.

Analyze Reports

Review LLM-as-a-Judge results in detail.