Output Formatter

Output Formatter policies format the final AI message output based on triggers. They are checked after the final AI message is generated, allowing you to transform or restructure the agent's response before it's returned to the user.

Overview

Output Formatters allow you to:

Transform responses: Reformat agent output using LLM-based instructions
Apply templates: Use structured formats (JSON, markdown) for consistent output
Direct replacement: Replace responses with predefined content when specific conditions are met
Conditional formatting: Apply formatting only when triggers match the agent's response

Usage

await agent.policies.add_output_formatter(
    name="Format as Summary",
    format_config="Format the response as a structured summary with:\n- A clear title\n- Key points as bullets",
    format_type="markdown",
    keywords=["summary", "result", "output"]
)

Format Types

Output Formatters support three format types:

Markdown Format

Uses LLM to reformat the response according to markdown instructions.

await agent.policies.add_output_formatter(
    name="Structured Report",
    format_config="""
    Format the response as a structured report with:
    - Executive Summary (2-3 sentences)
    - Key Findings (bullet points)
    - Recommendations (numbered list)
    - Next Steps (action items)
    """,
    format_type="markdown",
    keywords=["report", "analysis", "findings"]
)

JSON Schema Format

Uses LLM to reformat the response according to a JSON schema.

await agent.policies.add_output_formatter(
    name="Structured Data Output",
    format_config='''{
        "type": "object",
        "properties": {
            "status": {"type": "string"},
            "data": {"type": "array"},
            "summary": {"type": "string"}
        },
        "required": ["status", "data", "summary"]
    }''',
    format_type="json_schema",
    keywords=["data", "json", "structured"]
)

Direct Format

Replaces the response directly with the format_config string (no LLM call).

await agent.policies.add_output_formatter(
    name="Error Response Template",
    format_config="I apologize, but I encountered an error processing your request. Please try again or contact support.",
    format_type="direct",
    keywords=["error", "failed", "exception"]
)

Parameters

Prop

Type

Advanced Example with Natural Language Trigger

await agent.policies.add_output_formatter(
    name="Technical Documentation Format",
    format_config="""
    Format the response as technical documentation:
    1. Title (H1)
    2. Overview section
    3. Technical details in code blocks
    4. Examples section
    5. References
    """,
    format_type="markdown",
    natural_language_trigger=[
        "technical documentation",
        "API documentation",
        "code documentation",
        "developer guide"
    ],
    threshold=0.75
)

How It Works

Agent generates response: The agent completes its task and generates a final answer
Trigger matching: Output Formatter policies check if their triggers match the agent's response content
Formatting applied: If a match is found:
- Markdown/JSON Schema: LLM is called to reformat the response according to the instructions
- Direct: Response is replaced with the format_config string directly
Priority handling: If multiple formatters match, the one with highest priority is applied

Trigger Targets

Output Formatter triggers automatically target "agent_response" (the last AI message content). This ensures formatters are applied based on what the agent actually said, not the user's input.

Use Cases

Consistent formatting: Ensure all responses follow a specific structure
API responses: Format outputs as JSON for API integrations
Documentation: Transform responses into technical documentation format
Error handling: Replace error messages with user-friendly templates
Report generation: Structure analysis results as formatted reports
Data extraction: Convert unstructured responses into structured JSON

For detailed trigger configuration, see Triggers.

Output Formatter

On this page