dialectus-engine 0.3.0

Dialectus AI debate engine framework

Dialectus Engine

A Python library for orchestrating AI-powered debates with multi-provider model support.

Overview

The Dialectus Engine is a standalone Python library that provides core debate orchestration logic, including participant coordination, turn management, AI judge integration, and multi-provider model support. It's designed to be imported and used by other applications to build debate systems.

Components

• Core Engine (debate_engine/) - Main debate orchestration logic
• Models (models/) - AI model provider integrations (Ollama, OpenRouter)
• Configuration (config/) - System configuration management
• Judges (judges/) - AI judge implementations with ensemble support
• Formats (formats/) - Debate format definitions (Oxford, Parliamentary, Socratic, Public Forum)
• Moderation (moderation/) - Optional content safety system for debate topics

Installation

From PyPI

Using uv (recommended):

uv pip install dialectus-engine

Using pip:

pip install dialectus-engine

From Source

Using uv (recommended, faster):

# Clone the repository
git clone https://github.com/dialectus-ai/dialectus-engine.git
cd dialectus-engine

# Install in development mode with all dev dependencies
uv sync

# Or install without dev dependencies
uv pip install -e .

Using pip:

# Clone the repository
git clone https://github.com/dialectus-ai/dialectus-engine.git
cd dialectus-engine

# Install in development mode
pip install -e .

# Or install with dev dependencies
pip install -e ".[dev]"

As a Dependency

Add to your pyproject.toml:

[project]
dependencies = [
    "dialectus-engine>=0.1.0",
]

Or install directly from git:

# Using uv
uv pip install git+https://github.com/dialectus-ai/dialectus-engine.git@main

# Using pip
pip install git+https://github.com/dialectus-ai/dialectus-engine.git@main

Quick Start

import asyncio
from debate_engine.core import DebateEngine
from models.manager import ModelManager
from config.settings import AppConfig, ModelConfig

async def run_debate():
    # Load configuration
    config = AppConfig.from_json_file("debate_config.json")

    # Set up model manager
    model_manager = ModelManager()

    # Register models
    for model_id, model_config in config.models.items():
        model_manager.register_model(model_id, model_config)

    # Create debate engine
    engine = DebateEngine(
        config=config,
        model_manager=model_manager
    )

    # Run debate
    transcript = await engine.run_debate()
    print(transcript)

asyncio.run(run_debate())

Configuration

The engine uses debate_config.json for system configuration. To get started:

# Linux/Mac: Copy the example configuration
cp debate_config.example.json debate_config.json

# Windows (PowerShell):
# copy debate_config.example.json debate_config.json

# Edit with your settings and API keys
# Linux/Mac: nano debate_config.json
# Windows: notepad debate_config.json
# Or use your preferred editor (VS Code, vim, etc.)

Key configuration sections:

• Models: Define debate participants with provider, personality, and parameters
• Providers: Configure Ollama (local) and OpenRouter (cloud) settings
• Judging: Set evaluation criteria and judge models
• Debate: Default topic, format, and word limits
• Moderation (optional): Content safety for user-provided topics

For detailed configuration documentation, see CONFIG_GUIDE.md.

Development Workflows

Running Tests and Type Checking

Using uv (recommended):

# Run tests
uv run pytest

# Type check with Pyright
uv run pyright

# Lint with ruff
uv run ruff check .

# Format with ruff
uv run ruff format .

Using pip:

# Ensure dev dependencies are installed
pip install -e ".[dev]"

# Run tests
pytest

# Type check with Pyright
pyright

# Lint and format
ruff check .
ruff format .

Building Distribution

Using uv:

# Build wheel and sdist
uv build

# Install locally from wheel
uv pip install dist/dialectus_engine-*.whl

Using pip:

# Build wheel and sdist
python -m build

# Install locally
pip install dist/dialectus_engine-*.whl

Managing Dependencies

Using uv:

# Add a new dependency
# 1. Edit pyproject.toml [project.dependencies] section
# 2. Update lock file and sync environment:
uv lock && uv sync

# Upgrade all dependencies (within version constraints)
uv lock --upgrade

# Upgrade specific package
uv lock --upgrade-package httpx

# Add dev dependency
# 1. Edit pyproject.toml [project.optional-dependencies.dev]
# 2. Run:
uv sync

Using pip:

# Add a new dependency
# 1. Edit pyproject.toml dependencies
# 2. Reinstall:
pip install -e ".[dev]"

Why uv?

• 10-100x faster than pip for installs and resolution
• Reproducible builds via uv.lock (cross-platform, includes hashes)
• Python 3.14 ready - Takes advantage of free-threading for even better performance
• Single source of truth - Dependencies in pyproject.toml, lock file auto-generated
• Compatible - pip still works perfectly with pyproject.toml

Features

Multi-Provider Model Support

• Ollama: Local model management with hardware optimization
• OpenRouter: Cloud model access to 100+ models
• Async streaming: Chunk-by-chunk response generation
• Auto-discovery: Dynamic model listing from all configured providers
• Caching: In-memory cache with TTL for model metadata

Debate Formats

• Oxford: Classic opening/rebuttal/closing structure
• Parliamentary: British-style government vs. opposition
• Socratic: Question-driven dialogue format
• Public Forum: American high school debate style

AI Judge System

• LLM-based evaluation: Detailed criterion scoring
• Ensemble judging: Aggregate decisions from multiple judges
• Structured decisions: JSON-serializable judge results
• Configurable criteria: Logic, evidence, persuasiveness, etc.

Content Moderation (Optional)

• Multi-provider support: Ollama (local), OpenRouter, OpenAI moderation API
• Safety categories: Harassment, hate speech, violence, sexual content, dangerous activities
• Flexible deployment: Enable for production APIs, disable for trusted environments
• Graceful error handling: Provider-specific rate limit handling and retry logic

Architecture

Key architectural principles:

• Library-first: Designed to be imported by other applications
• Provider agnostic: Support for multiple AI model sources
• Async by default: All model interactions are async
• Type-safe: Strict Pyright configuration with modern type hints
• Pydantic everywhere: All config and data models use Pydantic v2
• Configurable: JSON-based configuration with validation

Technology Stack

• Python 3.13+ with modern type hints (X | None, list[T], dict[K, V])
• Pydantic v2 for data validation and settings management
• OpenAI SDK for OpenRouter API integration (streaming support)
• httpx for async HTTP requests (Ollama provider)
• asyncio for concurrent debate operations

Usage Examples

Listing Available Models

from models.manager import ModelManager

async def list_models():
    manager = ModelManager()
    models = await manager.get_all_models()
    for model_id, model_info in models.items():
        print(f"{model_id}: {model_info.description}")

Running a Custom Format

from formats.registry import format_registry

# Get available formats
formats = format_registry.list_formats()

# Load a specific format
oxford = format_registry.get_format("oxford")
phases = oxford.phases()

Ensemble Judging

from judges.factory import JudgeFactory

# Create judge with multiple models
config.judging.judge_models = ["openthinker:7b", "llama3.2:3b", "qwen2.5:3b"]
judge = JudgeFactory.create_judge(config.judging, model_manager)

# Get aggregated decision
decision = await judge.judge_debate(context)

Content Moderation

from dialectus.engine.moderation import ModerationManager, TopicRejectedError

# Create moderation manager
manager = ModerationManager(config.moderation, config.system)

# Validate user-provided topic
user_topic = "Should AI be regulated?"

try:
    result = await manager.moderate_topic(user_topic)
    # Topic is safe, proceed with debate
    print(f"Topic approved with confidence: {result.confidence}")
except TopicRejectedError as e:
    # Topic violates content policy
    print(f"Topic rejected: {e.reason}")
    print(f"Violated categories: {', '.join(e.categories)}")

For comprehensive moderation testing and setup instructions, see MODERATION_TESTING.md.