​
Quickstart Guide

This guide will have you running your first episode in under 5 minutes.

​
Prerequisites

  • An API key from the Labs Portal
  • Python 3.11+ (or any HTTP client)
  • An active subscription to at least one collection

​
Step 1: Get Your API Key

  1. Log in to the Labs Portal
  2. Navigate to API Keys
  3. Click Create New Key
  4. Copy and save your key (it won’t be shown again)

​
Step 2: Browse Available Scenarios

First, let’s see what collections you have access to: 
curl -X GET "$LABS_URL/api/v1/catalog" \
  -H "Authorization: Bearer $LABS_API_KEY"

​
Step 3: Create an Episode

Pick a scenario and create a training episode: 
# Create an episode
data = json.dumps({
    "collection_slug": "case-management",
    "scenario_slug": "back-injury-claim"
}).encode()

req = urllib.request.Request(
    f"{BASE_URL}/api/v1/episodes",
    data=data,
    headers={**headers, "Content-Type": "application/json"},
    method="POST"
)
with urllib.request.urlopen(req) as response:
    episode = json.loads(response.read())

episode_id = episode["episode_id"]
initial_obs = episode["initial_observation"]["content"]
tools = episode["tools"]

print(f"Episode ID: {episode_id}")
print(f"Initial observation: {initial_obs}")
print(f"Available tools: {len(tools)} tools")

​
Step 4: Submit Turns

Now submit your model’s responses and collect rewards: 
done = False
turn = 0

while not done:
    turn += 1

    # Your model generates a response (simplified for demo)
    model_response = f"Turn {turn}: Tell me more about your symptoms."

    # Submit the turn
    data = json.dumps({
        "messages": [
            {"role": "assistant", "content": model_response}
        ]
    }).encode()

    req = urllib.request.Request(
        f"{BASE_URL}/api/v1/episodes/{episode_id}/turns",
        data=data,
        headers={**headers, "Content-Type": "application/json"},
        method="POST"
    )
    with urllib.request.urlopen(req) as response:
        result = json.loads(response.read())

    # Extract results
    observation = result["observation"]["content"]
    reward = result["reward"]
    done = result["episode_complete"]

    print(f"Turn {turn}:")
    print(f"  Reward: {reward:.2f}")
    print(f"  Observation: {observation[:100]}...")

    if done:
        print(f"  Episode complete! Reason: {result.get('terminal_reason', 'unknown')}")

​
Step 5: Use Tool Calls (Optional)

Some scenarios have tools (artifacts to create, decisions to make). Tool calls are attached to the assistant message: 
# Submit a turn with a tool call (tool_calls attached to the message)
data = json.dumps({
    "messages": [
        {
            "role": "assistant",
            "content": "Based on our conversation, I'm creating an assessment.",
            "tool_calls": [
                {
                    "type": "artifact",
                    "name": "medical_assessment",
                    "input": {
                        "condition": "lower_back_strain",
                        "severity": "moderate",
                        "recommendation": "physical_therapy"
                    }
                }
            ]
        }
    ]
}).encode()

req = urllib.request.Request(
    f"{BASE_URL}/api/v1/episodes/{episode_id}/turns",
    data=data,
    headers={**headers, "Content-Type": "application/json"},
    method="POST"
)
with urllib.request.urlopen(req) as response:
    result = json.loads(response.read())

​
Complete Example

Here’s a full training loop (pseudocode - replace your_model with your actual model):

​
Next Steps

Integration Guides

Set up with TRL, OpenRLHF, or NeMo-Aligner

API Reference

Full endpoint documentation

Concepts

Understand episodes, rewards, and tools

Best Practices

Tips for effective training