API Reference

Complete API documentation for AutoRubric, organized by functional area.

Overview

AutoRubric exports 87 public items across the main module and autorubric.graders. This reference is organized into thematic chapters for easier navigation.

Quick Links

Chapter	Key Exports	Description
CANNOT_ASSESS Handling	CannotAssessConfig, CannotAssessStrategy	Configure handling of uncertain verdicts
Core Grading	Criterion, Rubric, EvaluationReport	Fundamental types for rubric-based evaluation
Dataset	DataItem, RubricDataset	Dataset management and serialization
Distribution Metrics	earth_movers_distance, ks_test	Statistical distribution comparisons
Ensemble	EnsembleEvaluationReport, JudgeVote	Multi-judge aggregation
Eval Runner	EvalRunner, evaluate()	Batch evaluation with checkpointing
Few-Shot	FewShotConfig, FewShotExample	Calibration with labeled examples
Graders	CriterionGrader, Grader, JudgeSpec	Grader implementations
Length Penalty	LengthPenalty, compute_length_penalty	Verbosity control
LLM Infrastructure	LLMConfig, LLMClient, generate()	LLM client and configuration
Meta-Rubric Evaluation	evaluate_rubric_standalone, evaluate_rubric_in_context	Assess rubric quality
Metrics	MetricsResult, compute_metrics	Agreement and correlation metrics
Multi-Choice	CriterionOption, MultiChoiceVerdict	Ordinal and nominal scales
Utilities	aggregate_token_usage, word_count	Helper functions

Import Patterns

Main Module

from autorubric import (
    # Core types
    Criterion,
    Rubric,
    CriterionVerdict,
    EvaluationReport,

    # LLM configuration
    LLMConfig,
    LLMClient,

    # Dataset
    DataItem,
    RubricDataset,

    # Evaluation
    EvalRunner,
    evaluate,

    # Metrics
    compute_metrics,
    MetricsResult,
)

Graders Module

from autorubric.graders import (
    CriterionGrader,
    Grader,
    JudgeSpec,
)

Meta Module

from autorubric.meta import (
    evaluate_rubric_standalone,
    evaluate_rubric_in_context,
    get_standalone_meta_rubric,
    get_in_context_meta_rubric,
)

Type Hierarchy

Grader (ABC)
└── CriterionGrader

EvaluationReport
└── EnsembleEvaluationReport

CriterionReport
└── EnsembleCriterionReport

MetricsResult
├── CriterionMetrics
├── OrdinalCriterionMetrics
├── NominalCriterionMetrics
└── JudgeMetrics

Architecture

Grading Flow

1. Rubric.grade() delegates to grader's grade() method
2. CriterionGrader treats single LLM as "ensemble of 1"
3. Makes concurrent LLM calls per criterion per judge via asyncio.gather()
4. Aggregates votes using configurable strategy
5. Returns EnsembleEvaluationReport (consistent interface)

Score Calculation

# Positive criteria: MET earns weight, UNMET earns 0
# Negative criteria: MET subtracts weight, UNMET contributes 0
weighted_sum = sum(verdict_value * criterion.weight for each criterion)
score = clamp(weighted_sum / total_positive_weight, 0, 1)  # if normalized
# Length penalty subtracted after base calculation

Conventions

• All graders return EnsembleEvaluationReport for consistent interface
• raw_score is always populated regardless of normalize setting
• Parse failures use conservative defaults (UNMET for positive, MET for negative weights)
• Filter error is not None results in training pipelines
• Rate limiting via LLMConfig.max_parallel_requests (per-provider semaphore)