LLM 输出结构化验证与自动修复框架设计师

You are an expert in designing structured output validation and auto-repair pipelines for LLM applications.

Given the following context:

• Application type: [e.g., chatbot, data extraction, code generation]
• Expected output format: [e.g., JSON, XML, Markdown, code]
• Validation schema: [paste JSON Schema, Pydantic model, or describe constraints]
• LLM provider: [e.g., OpenAI, Anthropic, local model]
• Error tolerance: [strict / lenient / best-effort]

Design a complete structured output validation and auto-repair framework:

1. Validation Layer Design

   • Schema definition strategy (JSON Schema, Pydantic, Zod, etc.)
   • Multi-level validation: syntax > schema > semantic > business rules
   • Streaming vs batch validation tradeoffs

2. Error Detection & Classification

   • Common failure modes (truncation, hallucinated fields, type mismatches)
   • Error severity scoring (recoverable vs fatal)
   • Confidence-based output filtering

3. Auto-Repair Strategies

   • Prompt-based retry with error context injection
   • Partial parse + targeted re-generation
   • Fallback chain: constrained decoding > regex extraction > re-prompt
   • Token budget management across retries

4. Implementation Blueprint

   • Middleware/interceptor architecture
   • Integration with existing frameworks (Instructor, Outlines, Guardrails)
   • Monitoring: success rate, retry count, latency overhead

5. Testing & Hardening

   • Adversarial test cases for edge cases
   • Regression suite for schema evolution
   • Performance benchmarks

Provide concrete code examples and architecture diagrams (as Mermaid) for the recommended approach.