Model Routing — Intelligent Model Selection Middleware

Automatically select the best LLM model and think mode based on task type, context size, and response confidence. Cut API costs 40-70% by routing simple tasks to fast models and complex tasks to capable ones.

How It Works

User message → Task Classifier → Model Router → Best Model → Response
                                     ↓
                          Low confidence? → Escalate to stronger model

Quick Start

# config.yaml
models:
  casual_chat:
    model: qwen3-14b
    think: false
  coding:
    model: qwen-coder
    think: true
  reasoning:
    model: deepseek-r1
    think: true
  long_context:
    model: glm-5.1
    think: false

from router import get_router

router = get_router()
result = await router.route("Write a Python web scraper")
# → Routes to qwen-coder with think=True

Features

• Task-type classification (coding, reasoning, chat, summarization)
• Per-model think mode configuration
• Confidence-based escalation (retry with stronger model)
• Context management and compaction at 55% threshold
• Hot-reload configuration (no restart needed)
• 83 tests passing

Cost Savings

License

MIT