LLM as judge

Score model outputs against a rubric. The same machinery as labeling, applied to evaluation.

A judge is a classifier whose input is a (prompt, response) pair and whose output is a score. Use int with ge and le so validated scores stay on scale.

from agno.agent import Agent
from agno.models.google import Gemini
from pydantic import BaseModel, Field


class Score(BaseModel):
    overall: int = Field(
        ..., ge=1, le=5, description="Overall quality, 5 is excellent"
    )


agent = Agent(
    model=Gemini(id="gemini-3.5-flash"),
    instructions=(
        "Score the response on overall quality from 1 (unusable) to 5 "
        "(excellent). Use the full scale. Reserve 5 for genuinely "
        "excellent responses."
    ),
    output_schema=Score,
)


def build_input(prompt: str, response: str) -> str:
    return f"Prompt:\n{prompt}\n\nResponse:\n{response}"


prompt = "Explain why the sky is blue, in one sentence."
result = agent.run(build_input(prompt, "It just is.")).content
# Score(overall=1)

Add a rationale

A free-text rationale makes the score auditable and surfaces rubric drift.

from pydantic import BaseModel, Field


class Score(BaseModel):
    overall: int = Field(..., ge=1, le=5, description="Overall quality")
    rationale: str = Field(..., description="Why this score, citing the response")

Multi-dimension rubric

Use one bounded field per rubric dimension and a separate field for the overall assessment.

from pydantic import BaseModel, Field


class RubricScore(BaseModel):
    correctness: int = Field(..., ge=1, le=5, description="Factually correct")
    completeness: int = Field(..., ge=1, le=5, description="Covers what was asked")
    clarity: int = Field(..., ge=1, le=5, description="Easy to follow")
    concision: int = Field(..., ge=1, le=5, description="No padding")
    overall: int = Field(..., ge=1, le=5, description="Holistic quality")

Picking the shape

You need	Schema
One quality number	int with ge=1, le=5
Number plus justification	Add a rationale field
Per-criterion breakdown	One bounded int field per dimension
Pairwise comparison	Preference data

Relationship to evals

This applies a classification schema to model outputs. When the judge is the deliverable, it lives here. When it scores a system under test, see Evals.

Next steps

Task	Guide
Rank two responses	Preference data
Reduce single-model bias	Quality pipeline

Developer Resources

• LLM-as-judge cookbook
• Evals