Output JSON Schema

Guide to defining structured output formats for Agent Nodes

Output JSON Schema ensures Agent Nodes generate responses in a consistent, structured JSON format. This is crucial for data parsing, conditional routing, and integration with other nodes. However, JSON Schema configuration alone is insufficient; clear prompt instructions are still required for optimal results.

Purpose

Problems Solved:

  • Inconsistent AI output that's difficult to parse
  • Need for structured data in Condition Nodes
  • Integration with APIs or other systems requiring fixed formats

Solution:

  • Define JSON schema for agent output
  • AI is constrained to follow the specified structure
  • Output is always valid JSON according to schema

How to Use

1. Set Output Format in Agent Node

In Agent Node Config Panel:

  1. Select Output Format → JSON
  2. Click "Add JSON Schema" or "Edit JSON Schema" button
  3. JSON Schema Editor will open

2. Define Schema

Two Editor Modes:

Builder Mode

Visual editor with drag-and-drop:

  • Click "Add property" for new fields
  • Set Name: property name (example: category, status)
  • Select Type: STR, NUM, BOOL, ENUM, OBJ, ARR
  • Fill Description: explain what AI should generate

Raw Mode

Edit JSON directly for complex schemas.

3. Property Types

String (STR)

{
  "name": "message",
  "type": "string",
  "description": "Summary text from analysis"
}

Number (NUM)

{
  "score": {
    "type": "number",
    "description": "Confidence score between 0-100"
  }
}

Boolean (BOOL)

{
  "is_urgent": {
    "type": "boolean",
    "description": "True if request is urgent"
  }
}

Enum (ENUM)

{
  "category": {
    "type": "string",
    "enum": ["A", "B", "C"],
    "description": "Classification result"
  }
}

Object (OBJ)

{
  "user_data": {
    "type": "object",
    "properties": {
      "name": { "type": "string" },
      "age": { "type": "number" }
    }
  }
}

Array (ARR)

{
  "items": {
    "type": "array",
    "items": {
      "type": "string"
    }
  }
}

Use Case Examples

1. Classification Agent

❌ Schema Only (Insufficient)

Schema:

{
  "type": "object",
  "properties": {
    "category": {
      "type": "string",
      "enum": ["A", "B", "C"]
    }
  }
}

Prompt:

Classify this request.

Problem: AI doesn't understand what categories A, B, C mean. Output may be random or inconsistent.

✅ Schema + Clear Instructions (Optimal)

Schema:

{
  "type": "object",
  "properties": {
    "category": {
      "type": "string",
      "enum": ["A", "B", "C"],
      "description": "Request classification result"
    }
  },
  "required": ["category"]
}

Instruction:

You are a precise **Request Classification Agent**. Your SOLE task is to analyze the user's input and categorize it into exactly one of three categories: "A", "B", or "C".

**CLASSIFICATION RULES:**

* **CATEGORY "A" (Information Retrieval)**
    * **Goal:** Fetching, reading, or searching for existing data.
    * **Triggers:** Search web, get stock price, read local file, list directory contents.

* **CATEGORY "B" (System Operations)**
    * **Goal:** Controlling the OS or generating Shell/CLI commands.
    * **Triggers:** Shutdown, restart, ping, check IP, make directory (mkdir), git commands.

* **CATEGORY "C" (General & Creative)**
    * **Goal:** Everything else.
    * **Triggers:** Coding questions, creative writing, translation, general chat, logic puzzles.

**FEW-SHOT EXAMPLES:**

**User:** "Search for today's BBCA stock price."
**Model:** { "category": "A" }

**User:** "Create a terminal command to delete the node_modules folder."
**Model:** { "category": "B" }

**User:** "Explain what a Black Hole is."
**Model:** { "category": "C" }

**User:** "Hello, how are you?"
**Model:** { "category": "C" }

Agent Output:

{
  "category": "A"
}

Use in Condition Node:

input.category == "A"

Note: JSON Schema only enforces output format, but doesn't guarantee correct content. Clear instructions with classification rules and few-shot examples ensure the AI understands the task and fills data accurately.

2. Data Extraction Agent

Schema:

{
  "type": "object",
  "properties": {
    "name": {
      "type": "string",
      "description": "Extracted person name"
    },
    "email": {
      "type": "string",
      "description": "Extracted email address"
    },
    "phone": {
      "type": "string",
      "description": "Extracted phone number"
    }
  },
  "required": ["name"]
}

Instruction:

You are a **Contact Information Extraction Agent**. Your task is to extract contact details from the provided text.

**EXTRACTION RULES:**
- Extract full name from the text
- Extract email address if present
- Extract phone number if present
- If information is not found, leave the field empty

**EXAMPLES:**

**User:** "Contact John Doe at john@example.com or call +1234567890"
**Model:** { "name": "John Doe", "email": "john@example.com", "phone": "+1234567890" }

**User:** "My name is Jane Smith, email: jane.smith@company.com"
**Model:** { "name": "Jane Smith", "email": "jane.smith@company.com", "phone": "" }

Agent Output:

{
  "name": "John Doe",
  "email": "john@example.com",
  "phone": "+1234567890"
}

3. Command Generator Agent

Schema:

{
  "type": "object",
  "properties": {
    "command": {
      "type": "string",
      "description": "Shell command to execute"
    },
    "explanation": {
      "type": "string",
      "description": "What the command does"
    }
  },
  "required": ["command", "explanation"]
}

Instruction:

You are a **System Command Generator**. Convert natural language requests into shell commands.

**GENERATION RULES:**
- Generate command for Windows PowerShell or Bash
- Provide clear explanation in the same language as user's request
- Use safe, standard commands

**EXAMPLES:**

**User:** "Check network connection"
**Model:** { "command": "ping google.com", "explanation": "Sends packets to google.com to check network connectivity" }

**User:** "Create folder named 'data_backup'"
**Model:** { "command": "mkdir data_backup", "explanation": "Creates a new folder named 'data_backup' in current directory" }

**User:** "List all files in current directory"
**Model:** { "command": "dir", "explanation": "Lists all files and directories in the current folder" }

Agent Output:

{
  "command": "ping google.com",
  "explanation": "Check network connectivity"
}

Use in Call Command Node:

input.command

4. Multi-Step Analysis

Schema:

{
  "type": "object",
  "properties": {
    "sentiment": {
      "type": "string",
      "enum": ["positive", "negative", "neutral"]
    },
    "score": {
      "type": "number",
      "description": "Confidence 0-100"
    },
    "keywords": {
      "type": "array",
      "items": {
        "type": "string"
      }
    }
  },
  "required": ["sentiment", "score"]
}

Instruction:

You are a **Sentiment Analysis Agent**. Analyze the emotional tone of the text.

**ANALYSIS RULES:**
- Classify sentiment as: positive, negative, or neutral
- Provide confidence score between 0-100
- Extract 3-5 key keywords that represent the sentiment

**EXAMPLES:**

**User:** "This product is excellent! Highly satisfied and would recommend."
**Model:** { "sentiment": "positive", "score": 95, "keywords": ["excellent", "satisfied", "recommend"] }

**User:** "The service was okay, nothing special."
**Model:** { "sentiment": "neutral", "score": 60, "keywords": ["okay", "nothing special"] }

**User:** "Very disappointed with the quality and customer support."
**Model:** { "sentiment": "negative", "score": 85, "keywords": ["disappointed", "quality", "support"] }

Agent Output:

{
  "sentiment": "positive",
  "score": 85,
  "keywords": ["excellent", "satisfied", "recommend"]
}

On this page

PurposeHow to Use1. Set Output Format in Agent Node2. Define SchemaBuilder ModeRaw Mode3. Property TypesUse Case Examples1. Classification Agent2. Data Extraction Agent3. Command Generator Agent4. Multi-Step Analysis