Generate Inference Result

Execute an agent and automatically capture traces, latency, and metadata

Overview

The generate() method is the recommended way to execute your agent when you want automatic trace collection, timing, and inference result creation. It handles the complete lifecycle:

1. Initializes trace collection
2. Executes your agent with timing measurement
3. Creates/updates the inference result with all metadata
4. Saves collected traces to the platform
5. Cleans up trace context (prevents memory leaks)

This is the recommended method for production use as it handles all trace lifecycle management automatically.

Agent Options

Simple

The quickest way to get started. Your function receives just the latest user message as a string.

def my_agent(user_message: str) -> str:
    # In a real scenario, call your model here
    return f"Your model output to: {user_message}"

Chat History

Use this when your agent needs the full conversation context. Your function receives the message list in the OpenAI format ({"role": "...", "content": "..."}).

def my_agent(messages: list[dict]) -> str:
    # messages follows the standard chat format:
    # [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}, ...]
    user_message = messages[-1]["content"]
    return f"Your model output to: {user_message}"

Structured

For full control over input and output — including optional usage tracking, cost tracking, and retrieval context for RAG evaluations.

def my_agent(input_data: AgentInput) -> AgentResponse:
    user_message = input_data.last_user_message_str()

    # Access structured input fields from the first user message's metadata
    # (e.g. when test case input is {"user_message": "hello", "chat_type": "support"})
    first_msg = input_data.messages[0] if input_data.messages else None
    chat_type = first_msg.metadata.get("chat_type") if first_msg and first_msg.metadata else None

    # In a real scenario, call your model here
    model_output = f"Your model output to: {user_message}"
    # Return AgentResponse with optional usage/cost tracking
    return AgentResponse(
        content=model_output,
        usage_info={"input_tokens": 100, "output_tokens": 50},
    )

All three signatures work with evaluations.run(), inference_results.generate(), and simulator.simulate(). Both sync and async functions are supported. The SDK auto-detects which signature you’re using from the type hint on the first parameter.

For the full list of fields available on AgentInput (including structured input access via message metadata), see the AgentInput reference.

Parameters

agent · AgentType · required

The agent function to execute. Supported signatures:

• Simple: (user_message: str) -> str — receives last user message
• Chat History: (messages: list[dict]) -> str — receives full chat history (OpenAI format)
• Structured: (input_data: AgentInput) -> AgentResponse — structured input/output. See the AgentInput reference for all available fields including structured input access via message metadata.

Both sync and async functions are supported. Use the structured signature when you need usage_info, cost_info, or retrieval_context.

session · Session · required

The session object to associate the inference result with. Obtain via galtea.sessions.create() or galtea.sessions.get().

input · str | dict

The input to send to the agent. Accepts a plain string or a dict with structured fields (e.g. {"user_message": "hello", "chat_type": "support"}). Stored directly in the inference result. If not provided and the session is test-based, the test case input is used automatically. Required for production/monitoring sessions. See Input Schema for how to define the expected field structure on your endpoint connection.

throw_if_empty_agent_response · bool

If True, raises a ValueError when the agent returns an empty response. Defaults to False.

context_data is automatically fetched from the session’s test case and passed to the agent via AgentInput.context_data. No need to provide it explicitly.

Returns

Returns an InferenceResult object with all captured data.

Example

# Define your agent function
def my_agent(user_message: str) -> str:
    return f"Response to: {user_message}"


inference_result = galtea.inference_results.generate(agent=my_agent, session=session, input="Generate something")

What Gets Captured

The generate() method automatically captures and saves:

Data	Source	Description
Input	input parameter	The user’s input (string or structured dict)
Output	AgentResponse.content	The agent’s response
Retrieval Context	AgentResponse.retrieval_context	Context used (for RAG). Stored as a RETRIEVER trace, not a readable field on the returned inference result.
Latency	Measured	End-to-end execution time in ms
Usage Info	AgentResponse.usage_info	Token counts (if provided)
Cost Info	AgentResponse.cost_info	Cost data (if provided)
Traces	@trace decorators	All traced operations

Providing Usage and Cost Information

Your agent can return usage and cost information in the AgentResponse:

@galtea.trace(name="main", type=TraceType.AGENT)
def my_agent_with_usage(input_data: AgentInput) -> AgentResponse:
    # Your agent logic...

    return AgentResponse(
        content="Response content",
        usage_info={
            "input_tokens": 150,
            "output_tokens": 75,
            "cache_read_input_tokens": 50,
        },
        cost_info={
            "cost_per_input_token": 0.00001,
            "cost_per_output_token": 0.00003,
            "cost_per_cache_read_input_token": 0.000001,
        },
    )

Comparison with Manual Approach

Feature	generate()	Manual with set_context()
Context initialization	Automatic	set_context()
Agent execution	Automatic	Manual call
Timing measurement	Automatic	Manual
Inference result	Auto-created	inference_results.create()
Trace export	Automatic	Automatic on clear_context()
Memory cleanup	Automatic	clear_context()
Error handling	Built-in	Manual try/finally

Use generate() for most use cases. Use manual set_context()/clear_context() only when you need to manage inference result creation separately or have complex multi-step workflows.

Error Handling

If the agent raises an exception, generate() ensures trace context is cleaned up:

def failing_agent(input_data: AgentInput) -> AgentResponse:
    raise ValueError("Agent failed intentionally")


try:
    result = galtea_client.inference_results.generate(agent=failing_agent, session=session_error, input="test")
except Exception as e:
    # Trace context is automatically cleaned up
    print(f"Agent failed: {e}")

Related Methods

• Create Inference Result - Manual creation
• Trace Service - Manual trace management
• Simulating Conversations - Multi-turn simulations