LLM Actions

Overview

LLM Actions in Agent Studio provide built-in capabilities to leverage large language model (LLM) functionalities directly within your Compound Actions and Conversational Processes. These actions enable tasks such as summarization, reasoning, classification, data extraction, and content generation, allowing you to build workflows without custom integrations.

This documentation focuses on two key LLM Actions: generate_text_action and generate_structured_value_action. These actions are designed to help you process unstructured data, generate insights, and structure outputs efficiently.

generate_text_action

Description

The generate_text_action invokes an LLM to produce free-form text output based on user-provided input. This action is ideal for tasks requiring natural language generation, such as summarizing documents, generating responses, or performing step-by-step reasoning. Use this action when you need unstructured text results, like drafting emails, explaining concepts, or brainstorming ideas.

Input Parameters

Field	Type	Required	Description
`system_prompt`	`string`	❌	Defines the model's behavior or instructions. For example, "Act as a helpful assistant that summarizes technical articles."
`user_input`	`string`	✅	The primary context or query for the LLM to process.
`model`	`string`	❌	Specifies the LLM model to use. See the Model Reference for available options. Defaults to `gpt-4o-mini-2024-07-18`
`temperature`	`number`	❌	Control the randomness of the output. Higher values will make the output more random, while lower values will make it more focused and deterministic
`reasoning_effort`	`string`	❌	Optional reasoning effort argument. Can be set to one of "minimal" (only `gpt-5` models), `"low"`, `"medium"`, or `"high"`. Must be left empty for non-reasoning models such as `gpt-4.1`.

Output

Field	Type	Description
`generated_output`	`string`	The LLM-generated text response.

Usage Examples

Here are practical examples demonstrating various LLM abilities. Each includes a sample request schema for integration into a Compound Action.

Example 1: Text Summarization

Summarize a lengthy article or user query into a concise overview.

- action:
    action_name: mw.generate_text_action
    input_args:
      system_prompt: '''Summarize the following text in 3-5 bullet points, focusing on key takeaways.'''
      user_input: data.article_content  # e.g., a long blog post fetched from an API
      model: '''gpt-4o-mini'''
			temperature: 0.7
    output_key: summary_output

system_prompt: '''Summarize the following text in 3-5 bullet points, focusing on key takeaways.'''
user_input: data.article_content  # e.g., a long blog post fetched from an API
model: '''gpt-4o-mini'''
temperature: 0.7

Example 2: Content Generation

Generate creative or instructional content, such as drafting a user email.

- action:
    action_name: mw.generate_text_action
    input_args:
      system_prompt: '''Write a professional email response based on the user's complaint.'''
      user_input: data.user_complaint  # e.g., "My order is delayed by two weeks."
    output_key: email_draft

system_prompt: '''Write a professional email response based on the user's complaint.'''
user_input: data.user_complaint  # e.g., "My order is delayed by two weeks."

Example 3: Step-by-Step Reasoning

Guide the LLM through logical reasoning for problem-solving.

- action:
    action_name: mw.generate_text_action
    input_args:
      system_prompt: '''Solve the problem step by step, explaining your reasoning.'''
      user_input: '''What is the next number in the sequence: 2, 4, 8, 16?'''
      reasoning_effort: '''high'''	
			model: '''gpt-5-2025-08-07'''
    output_key: reasoning_output

system_prompt: '''Solve the problem step by step, explaining your reasoning.'''
user_input: '''What is the next number in the sequence: 2, 4, 8, 16?'''
reasoning_effort: '''high'''	
model: '''gpt-5-2025-08-07'''

generate_structured_value_action

Description

The generate_structured_value_action calls an LLM to extract or generate data in a predefined structured format (JSON schema). This is particularly useful for classification, entity extraction, or transforming unstructured input into queryable data.

Apply this action for tasks where output consistency is critical, such as tagging content, extracting key-value pairs, or categorizing user inputs.

Input Parameters

Field	Type	Required	Description
`payload`	`object`	✅	The data or text to analyze.
`user_input`	`string`	❌	The primary context or query for the LLM to process.
`output_schema`	`object`	✅	JSON Schema defining the expected output structure
`system_prompt`	`string`	❌	Defines the model's behavior or instructions. For example, "Act as a helpful assistant that summarizes technical articles."
`model`	`string`	❌	Specifies the LLM model to use. Defaults to `"gpt-4o-mini-2024-07-18"`. IMPORTANT: This action is only compatible with `gpt-4o-mini-2024-07-18` and later and `gpt-4o-2024-08-06` and later.
`strict`	`boolean`	❌	Enforces schema adherence. Defaults to `false`.
`output_schema_name`	`string`	❌	LLM-facing name for the schema (defaults to `extracted_value`)
`output_schema_description`	`string`	❌	Description of the schema for the LLM.
`reasoning_effort`	`string`	❌	Optional reasoning effort argument. Can be set to one of "minimal" (only `gpt-5` models), `"low"`, `"medium"`, or `"high"`. Must be left empty for non-reasoning models such as `gpt-4.1`.

Output

Field	Type	Description
`generated_output`	`object`	Structured data matching the provided schema.

Usage Examples

Examples illustrate extraction, classification, and more. Include request schemas for easy implementation.

Example 1: Topic Classification (Existing Example, Expanded)

Classify a research abstract into predefined topics.

- action:
    action_name: mw.generate_structured_value_action
    input_args:
      payload: data.research_paper_abstract
      system_prompt: '''Given a research paper abstract and a list of topic options, output up to 5 topics that accurately apply to the paper.'''
      output_schema: >-
        {
            "type": "object",
            "properties": {
                "topic_tags": {
                    "type": "array",
                    "items": {
                      "type": "string",
                      "enum": data.topic_tag_options  # e.g., ["AI", "ML", "NLP"]
                    }
                }
            },
            "required": ["topic_tags"],
            "additionalProperties": false
        }
      strict: true
    output_key: classified_topics

payload: data.research_paper_abstract
system_prompt: '''Given a research paper abstract and a list of topic options, output up to 5 topics that accurately apply to the paper.'''
output_schema: >-
    {
        "type": "object",
        "properties": {
            "topic_tags": {
                "type": "array",
                "items": {
                    "type": "string",
                    "enum": data.topic_tag_options  # e.g., ["AI", "ML", "NLP"]
                }
            }
        },
        "required": ["topic_tags"],
        "additionalProperties": false
    }
strict: true

Expected Output

generated_output: {
  "topic_tags": ["LLM Capabilities", "Reinforcement Learning (RL)", "Reasoning"]
}

Example 2: Entity Extraction

Extract named entities like names, dates, and locations from text.

- action:
    action_name: mw.generate_structured_value_action
    input_args:
      payload: data.user_message  # e.g., "John Doe will arrive in New York on October 15, 2025."
      system_prompt: '''Extract entities such as persons, locations, and dates from the text.'''
      output_schema: >-
        {
            "type": "object",
            "properties": {
                "persons": {"type": "array", "items": {"type": "string"}},
                "locations": {"type": "array", "items": {"type": "string"}},
                "dates": {"type": "array", "items": {"type": "string"}}
            },
            "required": ["persons", "locations", "dates"],
            "additionalProperties": false
        }
      reasoning_effort: '''low'''
			model: '''gpt-5-2025-08-07'''
    output_key: extracted_entities

payload: data.user_message  # e.g., "John Doe will arrive in New York on October 15, 2025."
system_prompt: '''Extract entities such as persons, locations, and dates from the text.'''
output_schema: >-
    {
        "type": "object",
        "properties": {
            "persons": {"type": "array", "items": {"type": "string"}},
            "locations": {"type": "array", "items": {"type": "string"}},
            "dates": {"type": "array", "items": {"type": "string"}}
        },
        "required": ["persons", "locations", "dates"],
        "additionalProperties": false
    }
reasoning_effort: '''low'''
model: '''gpt-5-2025-08-07'''

Expected Output

generated_output: {
  "persons": ["John Doe"],
  "locations": ["New York"],
  "dates": ["October 15, 2025"]
}

Example 3: Sentiment Classification

Classify text sentiment with confidence scores.

- action:
    action_name: mw.generate_structured_value_action
    input_args:
      payload: data.customer_review
      system_prompt: '''Analyze the sentiment of the review and output the category with a confidence score.'''
      output_schema: >-
        {
            "type": "object",
            "properties": {
                "sentiment": {
                    "type": "string",
                    "enum": ["positive", "negative", "neutral"]
                },
                "confidence": {"type": "number"}
            },
            "required": ["sentiment", "confidence"],
            "additionalProperties": false
        }
    output_key: sentiment_analysis

payload: data.customer_review
system_prompt: '''Analyze the sentiment of the review and output the category with a confidence score.'''
output_schema: >-
    {
        "type": "object",
        "properties": {
            "sentiment": {
                "type": "string",
                "enum": ["positive", "negative", "neutral"]
            },
            "confidence": {"type": "number"}
        },
        "required": ["sentiment", "confidence"],
        "additionalProperties": false
    }

Expected Output

generated_output: {
  "sentiment": "positive",
  "confidence": 0.85
}

Model Reference

Model	Context Window	Max Output	Reasoning Effort Compatible	Best For	Availability	Live web search capable
`gpt-5-2025-08-07`	400k tokens	128k tokens	✅	Advanced multi-modal reasoning and long-form content creation	OpenAI direct only. Not available to Azure OpenAI.	❌
`gpt-5`	400k tokens	128k tokens	✅	Complex problem-solving with large datasets	OpenAI direct only. Not available to Azure OpenAI.	❌
`gpt-5-mini-2025-08-07`	400k tokens	128k tokens	✅	Efficient long-context analysis and summarization	OpenAI direct only. Not available to Azure OpenAI.	❌
`gpt-5-mini`	400k tokens	128k tokens	✅	Quick prototyping of AI agents with extended memory	OpenAI direct only. Not available to Azure OpenAI.	❌
`gpt-5-nano-2025-08-07`	400k tokens	128k tokens	✅	Processing of lengthy documents	OpenAI direct only. Not available to Azure OpenAI.	❌
`gpt-5-nano`	400k tokens	128k tokens	✅	Processing of lengthy documents	OpenAI direct only. Not available to Azure OpenAI.	❌
`o4-mini-2025-04-16`	200k tokens	100k tokens	✅	Step-by-step reasoning for technical documentation	OpenAI direct only. Not available to Azure OpenAI.	❌
`o4-mini`	200k tokens	100k tokens	✅	Optimized chain-of-thought for research tasks	OpenAI direct only. Not available to Azure OpenAI.	❌
`gpt-4.1-2025-04-14`	1M tokens	32k tokens	❌	Ultra-long document processing and archival analysis	Available to all commercial orgs.	❌
`gpt-4.1`	1M tokens	32k tokens	❌	Ultra-long document processing and archival analysis	OpenAI direct only. Not available to Azure OpenAI.	❌
`gpt-4.1-mini-2025-04-14`	1M tokens	32k tokens	❌	High-volume text extraction from large documents	Available to all commercial orgs.	❌
`gpt-4.1-mini`	1M tokens	32k tokens	❌	High-volume text extraction from large documents	OpenAI direct only. Not available to Azure OpenAI.	❌
`gpt-4.1-nano-2025-04-14`	1M tokens	32k tokens	32k tokens	Real-time processing of voluminous logs and transcripts	Available to all commercial orgs.	❌
`gpt-4.1-nano`	1M tokens	32k tokens	❌	Real-time processing of voluminous logs and transcripts	OpenAI direct only. Not available to Azure OpenAI.	❌
`o3-2025-04-16`	200k tokens	100k tokens	✅	Rigorous logical deduction	OpenAI direct only. Not available to Azure OpenAI.	❌
`o3`	200k tokens	100k tokens	✅	In-depth strategic planning	OpenAI direct only. Not available to Azure OpenAI.	❌
`o3-mini-2025-01-31`	200k tokens	100k tokens	✅	Educational tools for step-by-step learning paths	OpenAI direct only. Not available to Azure OpenAI.	❌
`o3-mini`	200k tokens	100k tokens	✅	Debugging complex workflows with explanatory outputs	OpenAI direct only. Not available to Azure OpenAI.	❌
`o1-2024-12-17`	200k tokens	100k tokens	✅	Mathematical proofs and logical puzzle solving	OpenAI direct only. Not available to Azure OpenAI.	❌
`o1`	200k tokens	100k tokens	✅	Creative brainstorming with structured ideation	OpenAI direct only. Not available to Azure OpenAI.	❌
`gpt-4o-2024-11-20`	128K tokens	16k tokens	❌	Content generation	OpenAI direct only. Not available to Azure OpenAI.	❌
`gpt-4o`	128K tokens	16k tokens	❌	Content generation	OpenAI direct only. Not available to Azure OpenAI.	❌
`gpt-4o-mini-2024-07-18`	128K tokens	16k tokens	❌	Quick content drafting and idea generation	OpenAI direct only. Not available to Azure OpenAI.	❌
`gpt-4o-mini`	128K tokens	16k tokens	❌	Quick content drafting and idea generation	OpenAI direct only. Not available to Azure OpenAI.	❌
`gpt-4o-search-preview`	128K tokens	16k tokens	❌	Real-time information retrieval and fact-checked responses	OpenAI direct only. Not available to Azure OpenAI.	✅

Updated 3 days ago