AutoEvaluator: LLM-Based Evaluation Framework

AutoEvaluator is a powerful Python library that accelerates LLM output quality control through automated evaluation. Using LLMs to evaluate LLMs, it provides a simple, transparent, and developer-friendly API to identify True Positives (TP), False Positives (FP), and False Negatives (FN) in generated content against ground truth.

🚀 Features

• Automated Evaluation: Compare LLM outputs against ground truth with precision
• Multi-Provider Support: Works with AWS Bedrock, OpenAI, Anthropic, and Google Gemini
• Comprehensive Metrics: Automatically calculates Precision, Recall, and F1 Score
• Async-First Design: Built for high-performance concurrent evaluations
• Structured Outputs: Leverages Instructor for type-safe, validated responses
• Sentence-Level Granularity: Evaluates claims at the sentence level for detailed insights

🔧 Installation

Requirements

• Python 3.9 or higher
• An API key for at least one supported LLM provider

Install via pip

pip install autoevaluator

Install from source

git clone https://github.com/yourusername/autoevaluator.git
cd autoevaluator
pip install -e .

⚡ Quick Start

import asyncio
from dotenv import load_dotenv
load_dotenv()  # Load env variables BEFORE importing autoevaluator
from autoevaluator import evaluate, get_instructor_client

async def main():
    # Setup client for your preferred provider
    client = get_instructor_client(provider="openai", model="gpt-4o-mini")
    
    # Define the claim to evaluate
    claim = "Feynman was born in 1918 in Malaysia"
    
    # Define the ground truth
    ground_truth = "Feynman was born in 1918 in America."
    
    # Evaluate the claim
    result = await evaluate(
        claim=claim,
        ground_truth=ground_truth,
        client=client,
        model_name="gpt-4o-mini"
    )
    
    print(result)

# Run the async function
asyncio.run(main())

Output:

{
    'TP': ['Feynman was born in 1918.'],
    'FP': ['Feynman was born in Malaysia.'],
    'FN': ['Feynman was born in America.'],
    'precision': 0.5,
    'recall': 0.5,
    'f1_score': 0.5
}

🔌 Supported Providers

AutoEvaluator supports multiple LLM providers out of the box:

Provider	Models	Environment Variables
AWS Bedrock	Claude Sonnet 4.5	AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_REGION
OpenAI	GPT-4o, GPT-4o-mini, etc.	OPENAI_API_KEY
Anthropic	Claude Sonnet 4, etc.	ANTHROPIC_API_KEY
Google Gemini	Gemini 2.0 Flash, etc.	GOOGLE_API_KEY

⚙️ Configuration

Environment Variables

Create a .env file in your project root:

# OpenAI
OPENAI_API_KEY=your_openai_api_key

# AWS Bedrock
AWS_ACCESS_KEY_ID=your_aws_access_key
AWS_SECRET_ACCESS_KEY=your_aws_secret_key
AWS_REGION=ap-southeast-1

# Anthropic
ANTHROPIC_API_KEY=your_anthropic_api_key

# Google Gemini
GOOGLE_API_KEY=your_google_api_key

Python Configuration

import os

# Set environment variables programmatically
os.environ["OPENAI_API_KEY"] = "your_openai_api_key"
os.environ["AWS_ACCESS_KEY_ID"] = "your_aws_access_key"
os.environ["AWS_SECRET_ACCESS_KEY"] = "your_aws_secret_key"

💡 Usage Examples

Example 1: Using OpenAI

import asyncio
from dotenv import load_dotenv
load_dotenv()  # Load env variables BEFORE importing autoevaluator
from autoevaluator import evaluate, get_instructor_client

async def evaluate_with_openai():
    client = get_instructor_client(provider="openai", model="gpt-4o-mini")
    
    claim = "The Earth is flat and the moon landing was in 1969."
    ground_truth = "The Earth is round. The moon landing was in 1969."
    
    result = await evaluate(claim, ground_truth, client=client, model_name="gpt-4o-mini")
    
    print(f"True Positives: {result['TP']}")
    print(f"False Positives: {result['FP']}")
    print(f"False Negatives: {result['FN']}")
    print(f"Precision: {result['precision']:.2f}")
    print(f"Recall: {result['recall']:.2f}")
    print(f"F1 Score: {result['f1_score']:.2f}")

asyncio.run(evaluate_with_openai())

Example 2: Using AWS Bedrock

import asyncio
from dotenv import load_dotenv
load_dotenv()  # Load env variables BEFORE importing autoevaluator
from autoevaluator import evaluate, get_instructor_client

async def evaluate_with_bedrock():
    client = get_instructor_client(provider="bedrock")
    
    claim = "Python was created by Guido van Rossum in 1991."
    ground_truth = "Python was created by Guido van Rossum in 1991."
    
    result = await evaluate(claim, ground_truth, client=client, model_name="bedrock-claude")
    return result

result = asyncio.run(evaluate_with_bedrock())
print(f"Perfect match! F1 Score: {result['f1_score']}")

Example 3: Using Anthropic

import asyncio
from dotenv import load_dotenv
load_dotenv()  # Load env variables BEFORE importing autoevaluator
from autoevaluator import evaluate, get_instructor_client

async def evaluate_with_anthropic():
    client = get_instructor_client(
        provider="anthropic",
        model="claude-sonnet-4-20250514"
    )
    
    claim = "Water boils at 100°C at sea level."
    ground_truth = "Water boils at 100°C at sea level."
    
    result = await evaluate(claim, ground_truth, client=client, model_name="claude-sonnet-4-20250514")
    return result

result = asyncio.run(evaluate_with_anthropic())

Example 4: Using Google Gemini

import asyncio
from dotenv import load_dotenv
load_dotenv()  # Load env variables BEFORE importing autoevaluator
from autoevaluator import evaluate, get_instructor_client

async def evaluate_with_gemini():
    client = get_instructor_client(
        provider="gemini",
        model="gemini-2.0-flash-exp"
    )
    
    claim = "The speed of light is approximately 300,000 km/s in a vacuum."
    ground_truth = "The speed of light is approximately 300,000 kilometers per second in a vacuum."
    
    result = await evaluate(
        claim=claim,
        ground_truth=ground_truth,
        client=client,
        model_name="gemini-2.0-flash-exp"
    )
    
    print(f"True Positives: {result['TP']}")
    print(f"False Positives: {result['FP']}")
    print(f"False Negatives: {result['FN']}")
    print(f"F1 Score: {result['f1_score']:.2f}")
    
    return result

result = asyncio.run(evaluate_with_gemini())

Example 5: Batch Evaluation

import asyncio
from dotenv import load_dotenv
load_dotenv()  # Load env variables BEFORE importing autoevaluator
from autoevaluator import evaluate, get_instructor_client

async def batch_evaluate():
    client = get_instructor_client(provider="openai", model="gpt-4o-mini")
    
    test_cases = [
        {
            "claim": "Einstein developed the theory of relativity.",
            "ground_truth": "Einstein developed the theory of relativity."
        },
        {
            "claim": "The capital of France is London.",
            "ground_truth": "The capital of France is Paris."
        },
        {
            "claim": "Water is composed of hydrogen and oxygen.",
            "ground_truth": "Water is composed of hydrogen and oxygen."
        }
    ]
    
    tasks = [
        evaluate(tc["claim"], tc["ground_truth"], client=client, model_name="gpt-4o-mini")
        for tc in test_cases
    ]
    
    results = await asyncio.gather(*tasks)
    
    for i, result in enumerate(results, 1):
        print(f"\n--- Test Case {i} ---")
        print(f"F1 Score: {result['f1_score']:.2f}")
        print(f"Precision: {result['precision']:.2f}")
        print(f"Recall: {result['recall']:.2f}")

asyncio.run(batch_evaluate())

📚 API Reference

evaluate()

Evaluates a claim against ground truth and returns detailed metrics.

async def evaluate(
    claim: str,
    ground_truth: str,
    client: instructor.AsyncInstructor,
    model_name: str = "gpt-4o-mini"
) -> Dict[str, Any]

Parameters:

• claim (str): The text to be evaluated
• ground_truth (str): The reference text to compare against
• client (instructor.AsyncInstructor): Instructor-wrapped async client
• model_name (str): Model identifier to use

Returns:

Dictionary containing:

• TP (List[str]): List of true positive sentences
• FP (List[str]): List of false positive sentences
• FN (List[str]): List of false negative sentences
• precision (float): Precision score (0.0 to 1.0)
• recall (float): Recall score (0.0 to 1.0)
• f1_score (float): F1 score (0.0 to 1.0)

get_instructor_client()

Creates an Instructor-wrapped client for the specified LLM provider.

def get_instructor_client(
    provider: Literal["bedrock", "openai", "anthropic", "gemini"] = "bedrock",
    model: Optional[str] = None,
    api_key: Optional[str] = None,
    mode: instructor.Mode = instructor.Mode.JSON,
    **kwargs
) -> instructor.AsyncInstructor

Parameters:

• provider (str): LLM provider to use ("bedrock", "openai", "anthropic", "gemini")
• model (Optional[str]): Model name (uses provider default if None)
• api_key (Optional[str]): API key (falls back to environment variables)
• mode (instructor.Mode): Instructor parsing mode
• **kwargs: Additional provider-specific arguments

Returns:

An Instructor-wrapped async client ready for use.

text_simplifier()

Breaks down complex text into simple, single-clause sentences.

async def text_simplifier(
    text: str,
    model_name: str,
    client: instructor.AsyncInstructor
) -> TextSimplify

🔍 How It Works

AutoEvaluator uses a sophisticated multi-step process to evaluate claims:

1. Text Simplification: Complex sentences are broken down into simple, atomic claims
2. Question Generation: Each simplified sentence is converted into a fact-checking question
3. Bidirectional Verification: Questions are checked against both the claim and ground truth
4. Classification: Sentences are classified as TP, FP, or FN based on verification results
5. Metrics Calculation: Precision, Recall, and F1 scores are computed from the classifications

Architecture

Input Claim & Ground Truth
         ↓
   Text Simplifier (breaks into atomic sentences)
         ↓
   Question Generator (creates fact-check questions)
         ↓
   Question Checker (verifies against ground truth)
         ↓
   Classification (TP/FP/FN assignment)
         ↓
   Metrics Calculation (Precision, Recall, F1)
         ↓
   Structured Output

🎯 Advanced Usage

Custom Text Simplification

from autoevaluator import text_simplifier, get_instructor_client

async def simplify_text():
    client = get_instructor_client(provider="openai")
    
    complex_text = """Although the weather was bad and it was raining heavily, 
                      we decided to go hiking because we had planned it for weeks."""
    
    result = await text_simplifier(
        text=complex_text,
        model_name="gpt-4o-mini",
        client=client
    )
    
    print("Simplified sentences:")
    for sentence in result.simplified_sentences:
        print(f"- {sentence}")

asyncio.run(simplify_text())

Using Provider-Specific Convenience Functions

from autoevaluator.client import (
    get_openai_instructor_client,
    get_bedrock_instructor_client,
    get_anthropic_instructor_client,
    get_gemini_instructor_client
)

# OpenAI
openai_client = get_openai_instructor_client(model="gpt-4o")

# Bedrock
bedrock_client = get_bedrock_instructor_client()

# Anthropic
anthropic_client = get_anthropic_instructor_client()

# Gemini
gemini_client = get_gemini_instructor_client(model="gemini-2.0-flash")

Error Handling

import asyncio
from dotenv import load_dotenv
load_dotenv()  # Load env variables BEFORE importing autoevaluator
from autoevaluator import evaluate, get_instructor_client

async def safe_evaluate():
    try:
        client = get_instructor_client(provider="openai")
        result = await evaluate(
            claim="Some claim",
            ground_truth="Some truth",
            client=client,
            model_name="gpt-4o-mini"
        )
        return result
    except ValueError as e:
        print(f"Configuration error: {e}")
    except Exception as e:
        print(f"Evaluation error: {e}")

asyncio.run(safe_evaluate())

📊 Performance Considerations

• Async by Default: All operations are asynchronous for better performance
• Batch Processing: Use asyncio.gather() for concurrent evaluations
• Rate Limiting: Be mindful of provider rate limits when running batch evaluations
• Caching: Consider caching results for repeated evaluations

🤝 Contributing

Contributions are welcome! Please follow these steps:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add amazing feature')
4. Push to the branch (git push origin feature/amazing-feature)
5. Open a Pull Request

📄 License

This project is licensed under the MIT License. See the LICENSE file for details.

🙏 Acknowledgments

• Built with Instructor for structured outputs
• Supports multiple LLM providers through unified interfaces
• Inspired by the need for automated, reliable LLM evaluation

📧 Contact

Darveen Vijayan

• LinkedIn: darveenvijayan
• Twitter: @DarveenVijayan
• Medium: LLMs: A Calculator for Words

📈 Changelog

Version 1.1.0

• Multi-provider support (OpenAI, Bedrock, Anthropic, Gemini)
• Async-first architecture
• Improved text simplification
• Enhanced error handling

---

Made with ❤️ by Darveen Vijayan