GuidesAI reliability
LLM as a judge: grading an AI’s answers automatically
A rubric rather than an overall score
| Criterion | Question asked to the judge |
|---|---|
| Faithfulness | Is every claim backed by the sources provided? |
| Relevance | Does the answer address the question asked? |
| Completeness | Is any expected information missing? |
| Tone | Does it follow the style guidelines? |
A structured grade
const res = await client.messages.create({
model: process.env.CLAUDE_MODEL!,
max_tokens: 1024,
tools: [{
name: 'grade',
description: 'Records the grade of the answer, criterion by criterion',
input_schema: {
type: 'object',
properties: {
faithful: { type: 'boolean' }, relevant: { type: 'boolean' }, complete: { type: 'boolean' },
justification: { type: 'string' },
},
required: ['faithful', 'relevant', 'complete', 'justification'],
},
}],
tool_choice: { type: 'tool', name: 'grade' },
messages: [{ role: 'user', content: rubric(question, sources, answer) }],
});
const grade = res.content.find((b) => b.type === 'tool_use')?.input;
Calibrating the judge
- 1Human sample50 answers graded by hand.
- 2Same sample, judgeCompare the two grades.
- 3Adjust the rubricUntil agreement is good.
Yes-or-no criteria give more stable grades than a 1 to 10 scale.
Use cases: evaluating a RAG system and test an AI chatbot. The technique: reliable JSON from an LLM.
Frequently asked questions
- Can the judge be the same model as the one being evaluated?
- It can, but a model tends to prefer its own answers. A different model, or a very factual rubric, limits that bias.
- Which biases should I know about?
- A preference for long answers, for the first answer shown in a comparison, and for its own style. Shuffle the order and score length separately.
- How do I know the judge is reliable?
- Have the same sample graded by humans and by the judge, and measure how much they agree. Adjust the rubric while agreement is low.