Evaluating Conversations

Learn how to evaluate multi-turn conversations using Galtea’s session-based workflow across test-based, past conversations, and production monitoring.

To accurately evaluate interactions within a dialogue, you can use Galtea’s session-based workflow. This approach allows you to log an entire conversation and then run evaluations on all of its turns at once.

Certain metrics are specifically designed for conversational analysis and require the full context:

• Role Adherence: Measures how well the AI stays within its defined role.
• Knowledge Retention: Assesses the model’s ability to remember and use information from previous turns.
• Conversation Completeness: Evaluates whether the conversation has reached a natural and informative conclusion.
• Conversation Relevancy: Assesses whether each turn in the conversation is relevant to the ongoing topic.

The Session-Based Workflow

Every evaluation follows the same two steps: first get the conversation into a Session, then evaluate that session.

1. Capture the conversation in a session

   A Session holds all the turns of one conversation. How the turns get there depends on your use case (simulated, ingested, or live). See step 1.

2. Evaluate the session

   Once the turns are logged, evaluate the whole conversation at once with evaluations.create(). See step 2.

1. Capture the conversation in a session

How turns get into a session depends on where the conversation comes from. Pick the use case that matches yours. Each one ends with a logged session you evaluate in step 2.

Test-based evaluation

Use this when you have test cases. Galtea’s Conversation Simulator drives a synthetic user against your own agent, so you provide an agent function. It requires test_case_id.

First, define an agent function that connects Galtea to your product:

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.

Then fetch a test case, create a session, and run the simulator with your agent:

# Fetch your test cases (created from a CSV of behavior tests)
test_cases = galtea_client.test_cases.list(test_id=behavior_test.id)
if not test_cases:
    raise ValueError("No test cases found")

# Take one test case (loop over `test_cases` to evaluate them all)
test_case = test_cases[0]

# Create a session for the test case, then let the simulator drive the
# conversation between a synthetic user and your agent function (`my_agent`)
session = galtea_client.sessions.create(version_id=version_id, test_case_id=test_case.id)
galtea_client.simulator.simulate(
    session_id=session.id,
    agent=my_agent,
    max_turns=test_case.max_iterations or 10,
)

See the full simulation workflow in Simulating Conversations.

Past conversations (offline ingestion)

Use this when the conversation already happened outside Galtea. Create a session (no test_case_id) and log all of its turns in a single batch.

# The conversation already happened: create a session (no test_case_id) and log every turn
session = galtea_client.sessions.create(
    version_id=version_id,
    custom_id="EXTERNAL_CONVERSATION_ID",  # optional: map to your own conversation ID
    is_production=True,  # set to True when these are real users
)

conversation_turns = [
    {"role": "user", "content": "What are some lower-risk investment strategies?"},
    {
        "role": "assistant",
        "content": "For lower-risk investments, consider diversified index funds, bonds, or Treasury securities.",
        "retrieval_context": "Low-risk investment options include index funds, government bonds, and Treasury securities.",
    },
    {"role": "user", "content": "With age, should the investment strategy change?"},
    {
        "role": "assistant",
        "content": "Yes, many advisors recommend shifting to more conservative investments as you approach retirement.",
        "retrieval_context": "Financial advisors typically recommend a more conservative asset allocation as investors near retirement age.",
    },
]

# Log all turns at once
galtea_client.inference_results.create_batch(session_id=session.id, conversation_turns=conversation_turns)

Monitoring (production)

Use this for real-time logging of user interactions from your live product. Create the session with is_production=True and log turns as they happen, or batch them at the end.

Log turns individually

# Create a production session and log each turn as it happens in your live app
session = galtea_client.sessions.create(version_id=version_id, is_production=True)


def your_product(user_input: str) -> str:
    return f"This is a simulated response to '{user_input}'"


def handle_turn(user_input: str) -> str:
    model_output = your_product(user_input)
    galtea_client.inference_results.create(session_id=session.id, input=user_input, output=model_output)
    return model_output


# Simulate production interactions
handle_turn("Hello!")
handle_turn("What services do you offer?")

Log turns in a batch

# Or, if you already have the full transcript, create the session and log all turns at once
session = galtea_client.sessions.create(version_id=version_id, is_production=True)

conversation_turns = [
    {"role": "user", "content": "What are some lower-risk investment strategies?"},
    {
        "role": "assistant",
        "content": "For lower-risk investments, consider diversified index funds, bonds, or Treasury securities.",
    },
]

galtea_client.inference_results.create_batch(session_id=session.id, conversation_turns=conversation_turns)

See the dedicated guide: Monitor Production Responses.

2. Evaluate the session

Whichever use case you picked, you now have a session with all its turns logged. Evaluate the whole conversation at once, either by passing Specification IDs so the API resolves their linked metrics automatically (recommended), or by listing metrics explicitly.

Using specification IDs

# Evaluate the whole conversation; metrics are resolved from the specifications automatically
galtea_client.evaluations.create(
    session_id=session.id,
    specification_ids=[conversation_spec.id],
)

Using metrics

# Evaluate the whole conversation by listing metrics explicitly
galtea_client.evaluations.create(
    session_id=session.id,
    metrics=[
        {"name": "Conversation Relevancy"},
        {"name": "Role Adherence"},
        {"name": "Knowledge Retention"},
    ],
)

For test-based evaluation, run this evaluation inside your loop so every simulated session is evaluated.

Custom Metrics with Full Conversation Access

When you use CustomScoreEvaluationMetric, your measure() method always receives an inference_results parameter containing InferenceResult objects. For session evaluations this includes all turns; for single inference result evaluations it contains one item. This enables conversation-level scoring (e.g., consistency checks, cross-turn analysis).

from galtea import CustomScoreEvaluationMetric, InferenceResult


class ConversationConsistency(CustomScoreEvaluationMetric):
    """Scores how consistently the assistant responds across all turns."""

    def __init__(self):
        super().__init__(name=metric_name)

    def measure(self, *args, inference_results: list[InferenceResult] | None = None, **kwargs) -> float:
        if not inference_results:
            return 0.0
        # Access the full conversation for cross-turn analysis
        assistant_outputs = [ir.actual_output for ir in inference_results if ir.actual_output]
        if len(assistant_outputs) < 2:
            return 1.0
        # Your custom logic here (e.g., check for contradictions across turns)
        return 0.9


galtea_client.evaluations.create(
    session_id=session.id,
    metrics=[
        {"name": "Role Adherence"},
        {"score": ConversationConsistency()},  # Custom multi-turn metric
    ],
)

Each InferenceResult object provides actual_input, actual_output, latency, index, and other fields. Retrieval context is not a field on the object; it is derived from the turn’s RETRIEVER trace spans. See the custom metrics tutorial for more on custom metrics.

Learn More

Session

A full conversation between a user and an AI system.

Inference Result

A single turn in a conversation between a user and the AI.

Evaluation

The assessment of an evaluation using a specific metric’s criteria

Conversation Simulator

Test your conversational AI with simulated multi-turn conversations