JSON Extraction Template

Table of Contents

Overview

Complexity: Simple (2 blocks) Use Case: Extract structured information from unstructured text as JSON

This template converts free-form text into structured JSON with title and description fields. Perfect for creating structured datasets from raw text content.

Pipeline Architecture

┌──────────────────┐     ┌──────────────────┐
│   Structured     │ ──► │       JSON       │
│    Generator     │     │    Validator     │
└──────────────────┘     └──────────────────┘

Input: content
  ↓
+ generated (title, description)
  ↓
+ valid, parsed_json

Blocks:

  1. StructuredGenerator - Extracts key information as structured JSON
  2. JSONValidatorBlock - Validates the JSON structure and required fields

Seed Format

Use the simplified content field in metadata:

{
  "repetitions": 3,
  "metadata": {
    "content": "Python is a high-level programming language known for readability and versatility. It's widely used in web development, data science, and automation."
  }
}

Multiple seeds example:

[
  {
    "repetitions": 2,
    "metadata": {
      "content": "Electric cars reduce emissions but require charging infrastructure."
    }
  },
  {
    "repetitions": 1,
    "metadata": {
      "content": "Machine learning algorithms learn patterns from data without explicit programming."
    }
  }
]

Output Format

Schema:

{
  "title": "string - concise title summarizing the content",
  "description": "string - detailed description of the content"
}

Example output:

{
  "title": "Introduction to Python",
  "description": "Python is a high-level programming language known for its readability and versatility, widely used in web development, data science, and automation."
}

Use Cases

Perfect for:

  • Creating structured datasets from raw text
  • Converting blog posts/articles to metadata
  • Extracting key information from descriptions
  • Building content catalogs with titles and summaries

Not ideal for:

  • Complex multi-field extraction (use custom blocks)
  • Binary classification tasks (use Text Classification template)
  • Question-answer pairs (use Q&A Generation template)

Customization

You can modify the template in lib/templates/json_generation.yaml:

Change output fields:

json_schema:
  type: object
  properties:
    headline: {type: string}
    summary: {type: string}
    category: {type: string}
  required: ["headline", "summary"]

Adjust LLM parameters:

config:
  temperature: 0.5  # Lower = more deterministic
  max_tokens: 512   # Limit response length

Customize the prompt:

user_prompt: "Extract the main topic and a brief summary from: {{ content }}"