agentic-eval-team

AgentCompany Label Evaluation System

A multi-agent system for evaluating and refining labels on text data using the smolagent framework. The system employs a manager-worker architecture where multiple agents collaborate to reach consensus on label quality.

Features

• Manager Agent: Analyzes dataset, assesses difficulty, creates worker agents with specific roles
• Worker Agents: Evaluate labels with diverse personas and perspectives
• Consensus Engine: Multiple strategies for resolving disagreements
• Dynamic Agent Generation: Manager decides how many workers to create and their specific roles
• Hybrid Resolution: Discussion first, then manager tie-break if needed
• Pydantic Validation: Type-safe input/output with Pydantic models
• Retry Logic: Automatic retry with exponential backoff for API calls
• Parallel Processing: Process multiple items concurrently with ThreadPoolExecutor

Supported Task Types

• Classification (sentiment, category, etc.)
• Translation Selection (choosing best translation)
• Keyword Extraction
• Custom label types (user-defined)

Architecture

┌─────────────────────────────────────────────────────────┐
│                    Input JSON File                      │
│  { "task_config": {...}, "items": [...] }               │
└─────────────────────────────────────────────────────────┘
                          │
                          ▼
┌─────────────────────────────────────────────────────────┐
│                  MANAGER AGENT (smolagent)              │
│  • Samples items to assess difficulty                  │
│  • Creates worker agents with specific roles           │
│  • Plans evaluation strategy                           │
│  • Tie-breaker when consensus fails                    │
└─────────────────────────────────────────────────────────┘
                          │
          ┌───────────────┼───────────────┐
          ▼               ▼               ▼
    ┌───────────┐   ┌───────────┐   ┌───────────┐
    │  Worker 1 │   │  Worker 2 │   │  Worker N │
    │ (strict)  │◄──┤ (creative)│◄──┤ (domain)  │
    └───────────┘   └───────────┘   └───────────┘
          │               │               │
          └───────────────┴───────────────┘
                          │
                          ▼
┌─────────────────────────────────────────────────────────┐
│                  CONSENSUS ENGINE                       │
│  • Check consensus threshold (configurable, default 60%)│
│  • Resolve via majority vote, union, or discussion     │
│  • Manager tie-break if still disagreement             │
└─────────────────────────────────────────────────────────┘
                          │
                          ▼
┌─────────────────────────────────────────────────────────┐
│                   Output JSON File                      │
│  [{ "id": "...", "labels": {...}, "evaluation": {...} }]│
└─────────────────────────────────────────────────────────┘

Installation

From PyPI

pip install agentic-eval-team

From source

pip install -e .

Requirements

• Python 3.10+
• smolagents
• openai
• pydantic
• tqdm

Quick Start

1. Prepare Input File

Create an input JSON file with task_config and items:

{
  "task_config": {
    "description": "Evaluate whether the assigned categories and sentiments are accurate",
    "evaluation_criteria": ["accuracy", "consistency", "completeness"],
    "consensus_strategy": "discussion_then_vote",
    "consensus_threshold": 0.6,
    "max_discussion_rounds": 2
  },
  "items": [
    {
      "id": "sample_001",
      "text": "The recent advancement in AI has revolutionized healthcare diagnostics.",
      "labels": {
        "category": "technology",
        "sentiment": "positive"
      }
    }
  ]
}

2. Run Evaluation

# Using the CLI command (after installation)
agentic-eval input.json -o output.json --mock

# Or using Python module
python -m agentic_eval_team input.json -o output.json --mock

# With a real model
agentic-eval input.json -o output.json --endpoint http://localhost:8000/v1 --model llama-3.1-8b

Configuration

Command Line Arguments

Task Config Options

Consensus Strategies

Project Structure

agentic-eval-team/
├── agentic_eval_team/           # Main package
│   ├── __init__.py
│   ├── __main__.py            # CLI entry point
│   ├── config.py              # Configuration
│   ├── models/
│   │   ├── manager.py         # Manager agent
│   │   ├── worker.py          # Worker agents (with retry logic)
│   │   ├── tools.py           # Manager tools
│   │   ├── schema.py          # Pydantic models
│   │   └── mock_model.py      # Mock model for testing
│   ├── consensus/
│   │   ├── engine.py          # Consensus orchestration
│   │   └── strategies.py      # Resolution strategies
│   ├── evaluation/
│   │   └── runner.py          # Parallel processing runner
│   ├── tasks/
│   │   ├── router.py          # Task type detection
│   │   └── prompts.py          # Prompt templates
│   └── utils/
│       ├── io.py              # JSON I/O utilities
│       ├── retry.py           # Retry decorator
│       └── errors.py          # Custom exceptions
├── samples/
│   └── input_sample.json      # Sample input
├── tests/
│   └── test_core.py           # Unit tests
├── pyproject.toml            # Package configuration
└── README.md

Testing

python -m unittest discover tests -v

Mock Testing

Use --mock flag to test without a running LLM server:

agentic-eval samples/input_sample.json -o output.json --mock --parallel 2

Key Improvements

Retry Logic

Worker agents automatically retry failed API calls with exponential backoff:

• Max retries: 3 (configurable)
• Initial delay: 1 second
• Backoff factor: 2x per retry
• Graceful fallback to error response if all retries fail

Parallel Processing

Items can be processed in parallel using ThreadPoolExecutor:

agentic-eval input.json --parallel 4  # Process 4 items concurrently

Configurable Consensus Threshold

The consensus_threshold in task_config controls when consensus is reached:

• 0.6 = 60% agreement required
• 1.0 = full consensus required

How It Works

1. Manager Analysis

The manager samples items from the dataset and assesses:

• Difficulty level (low/medium/high)
• Recommended strategies
• Number of discussion rounds

2. Agent Generation

Based on difficulty and task, workers are created with specific roles:

• strict_evaluator: Focuses on accuracy and correctness
• creative_reviewer: Looks for edge cases and alternatives
• domain_expert: Checks technical/domain accuracy (for high difficulty)
• lenient_reviewer: Evaluates practical acceptability (for high difficulty)

3. Evaluation Loop

For each item:

1. All workers independently evaluate the labels
2. Consensus is checked against the threshold
3. If no consensus, all workers discuss and re-evaluate
4. After max rounds, manager tie-breaks if still disagreement

4. Output

Each item in the output includes:

• Original id and text
• Final labels (possibly modified)
• evaluation_summary with consensus info

Example Output

{
  "id": "sample_001",
  "text": "The recent advancement in AI...",
  "labels": {
    "category": "technology",
    "sentiment": "positive"
  },
  "evaluation_summary": {
    "consensus": "correct",
    "reasoning": "Threshold reached: correct (2/2 >= 0.6)",
    "resolved_by": "workers",
    "worker_count": 2,
    "discussion_rounds": 0
  }
}