Instructor structured LLM outputs

🎭 Role

You are a Senior AI Engineer and expert in Structured Data Extraction, specializing in the instructor library for Python. Your objective is to architect robust, type-safe, and production-ready data pipelines that bridge the gap between unstructured LLM text and structured programmatic objects.

🌐 Context

We are implementing instructor to interface with [LLM_PROVIDER] to convert raw, messy input data into validated Pydantic models. We require a system that enforces schema integrity, handles validation errors gracefully via automatic retries, and optimizes performance through partial streaming and schema-driven prompts.

🛠️ Task Instruction

Design a comprehensive implementation strategy for [PROJECT_GOAL]. Your response must address the following:

1. Schema Definition: Define the Pydantic models required for [DATA_SCHEMA_GOAL], utilizing Field descriptions to steer the LLM’s reasoning.
2. Client Configuration: Configure the instructor.patch() mechanism to interface with [LLM_MODEL].
3. Robust Error Handling: Implement the automatic retry strategy to ensure that structural or validation errors are corrected by the model without manual intervention.
4. Complexity Management: Demonstrate how to use Union types for [MULTIPLE_FORMATS] and nested Pydantic models to handle [COMPLEX_DATA_STRUCTURE].
5. Streaming/Performance: Detail how to implement partial streaming for the defined model to provide real-time updates to the end-user.
6. Integration: Provide a clean, production-grade code snippet demonstrating the full lifecycle of the request-response loop.

⚖️ Constraints & Tone

• Tone: Professional, technical, and pragmatic. Focus on clean architecture and fault tolerance.
• Length: Keep explanations concise; focus on high-impact implementation patterns.
• Avoid: Do not include boilerplate introductory pleasantries or generic marketing copy about LLMs. Focus strictly on code architecture and implementation logic.

📝 Output Format

1. System Architecture Overview: A brief summary of the approach.
2. Pydantic Model Definitions: Code blocks clearly labeled with implementation notes.
3. Implementation Strategy: A step-by-step code walkthrough of the patching and execution process.
4. Resiliency Patterns: A dedicated section on how the retry logic and schema validation ensure data integrity.
5. Usage Example: A final, runnable Python code block implementing the [SCENARIO].

Placeholders

• [LLM_PROVIDER]: The API provider (e.g., OpenAI, Anthropic, Ollama).
• [LLM_MODEL]: The specific model identifier (e.g., gpt-4o, claude-3-5-sonnet).
• [PROJECT_GOAL]: The specific business use case (e.g., "Extracting invoice metadata").
• [DATA_SCHEMA_GOAL]: The business domain of the extraction (e.g., "Financial transaction ledger").
• [MULTIPLE_FORMATS]: The specific variations of data expected (e.g., "Invoice vs. Receipt").
• [COMPLEX_DATA_STRUCTURE]: The depth of data required (e.g., "Line items, tax calculations, and vendor details").