TrueLocale Consensus Engine

TrueLocale is an open-source, multi-agent consensus engine designed for high-accuracy translations. It orchestrates multiple LLMs (Translator, Reviewer, Judge, Arbitrator) to autonomously translate, verify, critique, and refine content until a consensus quality threshold is met.

🚀 Features

• Multi-Agent Workflow: Not just a wrapper. Uses a debate protocol between Agents to ensure accuracy.
• Role-Based Logic: separate personas for Translator (Proposes), Reviewer (Critiques), and Arbitrator (Resolves disputes).
• Provider Agnostic: Works with OpenAI, Anthropic, Google Gemini via the Vercel AI SDK.
• Configurable Thresholds: Define your own quality bars for "Auto-Approval" vs "Human Review".
• Stateless: Designed to be used in serverless environments (Next.js API Routes, Cloud Functions).

📦 Installation

npm install @evnola/truelocale
# or
yarn add @evnola/truelocale

🛠️ Usage

1. Initialize the Engine

You need to provide API keys and model configurations.

import { ConsensusEngine, EngineOptions } from '@evnola/truelocale';

const options: EngineOptions = {
  llmConfig: {
    TRANSLATOR: { provider: 'openai', model: 'gpt-4o' },
    REVIEWER: { provider: 'anthropic', model: 'claude-3-5-sonnet-20240620' },
    ARBITRATOR: { provider: 'openai', model: 'gpt-4o' },
    JUDGE: { provider: 'google', model: 'gemini-1.5-pro' },
    'translator-evaluator': { provider: 'openai', model: 'gpt-4o' }
  },
  thresholds: {
    NO_DISPUTE: 0.40,      // Below this score, keep original
    ARBITRATOR_UPPER: 0.85 // Above this, trust the Reviewer
  }
};

// Ensure API Keys are in process.env (OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.)
const engine = new ConsensusEngine(options);

2. Process a Segment

const result = await engine.processSegment({
  id: 'segment-1',
  index: 0,
  source: 'Hello world, this is a test.',
  currentTranslation: '',
  status: 'PENDING',
  history: [],
}, 'tr'); // Target Language: Turkish

console.log('Final Translation:', result.currentTranslation);
console.log('Status:', result.status);

🧠 Architecture

The engine follows a loop:

1. Translator proposes a draft.
2. Reviewer critiques the draft based on linguistic rules and context.
3. Translator evaluates the critique.
   • Accepts: Updates translation.
   • Rejects: Escalates to Arbitrator.
4. Arbitrator (if needed) rules on the dispute.

🤝 Contributing

We welcome contributions! Please see CONTRIBUTING.md for details.

📄 License

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