多模型智能路由规则自然语言生成器

You are an AI infrastructure architect specializing in multi-model routing and cost optimization. Help me design an intelligent model routing configuration. ## My Setup - **Available models**: [list your models, e.g., GPT-4o, Claude Sonnet, Gemini Flash, Llama 3.1 70B local] - **Monthly budget**: $[amount] - **Primary use cases**: [customer support / coding assistant / content generation / data analysis / etc.] - **Latency requirements**: [real-time < 2s / near-real-time < 5s / batch OK] - **Quality priority**: [accuracy-first / speed-first / cost-first / balanced] ## Generate the following: ### 1. Task Classification Rules Create a decision tree that classifies incoming requests into complexity tiers: - **Tier 1 (Simple)**: Pattern matching criteria → cheapest model - **Tier 2 (Medium)**: Pattern matching criteria → mid-tier model - **Tier 3 (Complex)**: Pattern matching criteria → premium model - **Tier 4 (Critical)**: Pattern matching criteria → best model + verification Include concrete examples for each tier. ### 2. Routing Configuration Generate a JSON/YAML configuration file compatible with common routing frameworks (LiteLLM, OpenRouter, or custom) including: - Model priority lists with fallback chains - Rate limits per model - Cost caps and alerts - Retry policies - Timeout settings ### 3. Quality Gates - Confidence score thresholds for auto-escalation - Output validation rules (format, length, safety) - A/B testing configuration for model comparison ### 4. Cost Monitoring Rules - Daily/weekly budget allocation - Alert thresholds (50%, 80%, 95% of budget) - Automatic downgrade rules when approaching limits - Cost-per-request tracking dimensions ### 5. Estimated Cost Breakdown Based on the use cases described, estimate: - Monthly token consumption per tier - Cost per model - Total monthly cost - Potential savings vs. single-model approach Output everything as production-ready configuration files with inline comments.