Skip to Content
Getting StartedQuick Start

Quick Start

Get from zero to your first prompt optimization in under 5 minutes.

Install the SDK

pip install kaizen-sdk

Configure your environment

The SDK reads two environment variables. Set them via export or a .env file:

export KAIZEN_API_KEY="kaizen_your-api-key"
export KAIZEN_BASE_URL="http://localhost:8000"

You can also pass these directly to the client constructor:

from kaizen_sdk import CTClient
 
client = CTClient(api_key="kaizen_...", base_url="http://localhost:8000")

Create a task and log feedback

A task represents a named unit of work in your LLM application (e.g., “summarize_ticket”). Once created, log feedback entries that pair inputs, outputs, and quality scores.

from kaizen_sdk import CTClient
 
client = CTClient()
 
# Create a task
task = client.create_task(
    name="summarize_ticket",
    description="Summarize support tickets into one paragraph",
    feedback_threshold=50
)
 
# Log feedback from your application
result = client.log_feedback(
    task_id=task.id,
    inputs={"text": "Server is down, users cannot log in..."},
    output="The server is experiencing an outage affecting user authentication.",
    score=0.85,
    source="sdk"
)

Auto-optimization triggers after 50 feedback entries by default. You can configure this per task via the feedback_threshold parameter when creating a task.

Check your progress

List your tasks to see how close each one is to the optimization threshold:

tasks = client.list_tasks()
for t in tasks:
    print(f"{t.name}: {t.feedback_count}/{t.feedback_threshold} "
          f"({t.threshold_progress:.0%} to optimization)")

Get your optimized prompt

Once optimization completes, retrieve the active prompt for your task:

prompt = client.get_prompt("summarize_ticket")
print(prompt.prompt_text)
print(f"Eval score: {prompt.eval_score}")
print(f"Version: {prompt.version_number}")

The prompt is cached locally for 5 minutes (configurable via cache_ttl). After a new optimization completes and is activated, the SDK automatically picks up the new version.

What happens under the hood

When your feedback count reaches the threshold:

  1. A Celery worker picks up the optimization job
  2. DSPy MIPROv2 runs with your feedback data (20/80 train/validation split)
  3. The best prompt is saved as a draft version
  4. An auto-PR is created in your Git repository with before/after scores
  5. You review and merge the PR, or activate the prompt via the API

Next steps

Last updated on
IntroductionSelf-Hosting