Structured Output Component

The Structured Output component transforms LLM responses into structured data formats based on predefined schemas. It ensures consistent output formatting, validates against schemas, and supports multiple output generation.

Structured Output workflow and architecture

Configuration Parameters

Required Parameters

llm: LLM configuration
- model: LLM model name
- temperature: Generation temperature
- maxTokens: Token limit
inputMessage: Input prompt or message
schemaName: Name of output schema to use
outputSchema: Schema definition
- fields: Field definitions
- types: Data types
- validations: Validation rules

Optional Parameters

generateMultiple: Generate multiple outputs (default: false)

Output Format

{
  "structured_output": {
    "data": Array<{
      [key: string]: any  // Matches schema definition
    }>,
    "schema_used": {
      "name": string,
      "version": string,
      "fields": Object
    }
  },
  "metadata": {
    "generation_time": number,
    "model_used": string,
    "temperature": number,
    "tokens_used": number
  },
  "validation": {
    "success": boolean,
    "errors": Array<{
      "field": string,
      "error": string,
      "value": any
    }>,
    "warnings": Array<string>
  },
  "statistics": {
    "outputs_generated": number,
    "fields_populated": number,
    "validation_pass_rate": number
  },
  "processing": {
    "start_time": string,
    "end_time": string,
    "duration": number,
    "retries": number
  }
}

Features

Schema validation
Multiple output support
Type checking
Field validation
Error handling
Performance metrics
Retry logic
Statistics tracking

Note: Define clear schemas with proper validation rules. Handle edge cases in field validations.

Tip: Use appropriate temperature settings for consistent output. Implement retry logic for failed generations.

Example Usage

const structuredOutput = new StructuredOutput();

// Single output generation
const result1 = await structuredOutput.generate({
  llm: {
    model: "gpt-4",
    temperature: 0.7,
    maxTokens: 1000
  },
  inputMessage: "Analyze the sentiment of this review",
  schemaName: "sentimentAnalysis",
  outputSchema: {
    fields: {
      sentiment: { type: "string", enum: ["positive", "negative", "neutral"] },
      confidence: { type: "number", min: 0, max: 1 },
      keywords: { type: "array", items: "string" },
      summary: { type: "string", maxLength: 200 }
    }
  }
});

// Multiple output generation
const result2 = await structuredOutput.generate({
  llm: {
    model: "gpt-4",
    temperature: 0.9,
    maxTokens: 2000
  },
  inputMessage: "Generate product names and descriptions",
  schemaName: "productIdeas",
  outputSchema: {
    fields: {
      name: { type: "string", maxLength: 50 },
      description: { type: "string", maxLength: 200 },
      targetMarket: { type: "string" },
      priceRange: { 
        type: "object",
        properties: {
          min: { type: "number" },
          max: { type: "number" }
        }
      }
    }
  },
  generateMultiple: true,
  count: 3
});

Common Schemas:

// Sentiment Analysis
{
  "sentiment": "string",
  "confidence": "number",
  "keywords": "string[]"
}

// Product Information
{
  "name": "string",
  "price": "number",
  "features": "string[]"
}

// Person Details
{
  "name": "string",
  "age": "number",
  "contact": {
    "email": "string",
    "phone": "string"
  }
}