Deployment

After a prompt optimization experiment completes, deploy the best prompt to make it available at runtime. Your application fetches the deployed prompt from Promptic instead of hardcoding it.

Deployment applies to prompt components. Tool Optimization produces descriptions you copy back to your stack.

Before deploying

Confirm that the selected iteration:

meets the held-out quality threshold;
has been reviewed on case-level failures;
uses an allowed model and provider;
has acceptable latency and projected cost;
contains no unintended instructions or exposed dataset content.

Deploy an experiment

Using the SDK

Python

from promptic_sdk import PrompticClient

with PrompticClient() as client:
    client.deploy(
        component_id="comp_...",
        experiment_id="exp_...",
    )

Using the CLI

Terminal

promptic deployments deploy <component-id> --experiment <experiment-id>

Using the API

Terminal

curl -X POST https://promptic.eu/api/v1/components/COMPONENT_ID/deployment \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"experimentId": "EXPERIMENT_ID"}'

Fetch the deployed prompt

Use get_deployed_prompt to retrieve the currently deployed prompt at runtime:

Python

from promptic_sdk import PrompticClient

with PrompticClient() as client:
    prompt = client.get_deployed_prompt("comp_...")

    if prompt:
        print(prompt["prompt"])          # The prompt text
        print(prompt["model"])           # e.g., "gpt-5.6-luna"
        print(prompt["provider"])        # e.g., "openai"
        print(prompt["score"])           # Best iteration score

Multi-message prompts

For prompts with system/user/assistant messages:

Python

prompt = client.get_deployed_prompt("comp_...")

if prompt and prompt["promptFormat"] == "multi_message":
    for msg in prompt["promptMessages"]:
        print(f"{msg['role']}: {msg['content']}")

Using the deployed prompt with OpenAI

Python

from openai import OpenAI
from promptic_sdk import PrompticClient

openai_client = OpenAI()

with PrompticClient() as client:
    deployed = client.get_deployed_prompt("comp_...")

    if deployed:
        messages = [{"role": m["role"], "content": m["content"]}
                    for m in deployed["promptMessages"]]
        messages.append({"role": "user", "content": user_input})

        response = openai_client.chat.completions.create(
            model=deployed["model"],
            messages=messages,
        )

Check deployment status

Python

with PrompticClient() as client:
    deployment = client.get_deployment("comp_...")

    if deployment:
        print(f"Experiment: {deployment['experimentId']}")
    else:
        print("No active deployment")

Remove a deployment

Python

with PrompticClient() as client:
    client.undeploy("comp_...")

Or via CLI:

Terminal

promptic deployments undeploy <component-id>

Runtime behavior

Treat the deployed prompt endpoint as configuration infrastructure:

cache only as long as your rollout can tolerate stale configuration;
define application behavior when no deployment exists or the fetch fails;
keep a known-good local fallback for critical paths;
record the component and experiment IDs with application release metadata;
validate variable interpolation before sending the prompt to a model.

Deploying another completed experiment replaces the component’s active deployment. To roll back, deploy the previous known-good experiment again. Removing a deployment makes the runtime endpoint return no active prompt; it does not delete the experiment.

On this page