Prompt Versioning: Create, Label, and Serve Prompt Versions

import os
from fi.prompt import Prompt
from fi.prompt.types import PromptTemplate, SystemMessage, UserMessage, ModelConfig

prompt_client = Prompt(
    template=PromptTemplate(
        name="support-response",
        messages=[
            SystemMessage(
                content="You are a helpful customer support agent for TechStore. "
                        "Answer the customer's question clearly and professionally."
            ),
            UserMessage(
                content="Customer question: {{question}}"
            ),
        ],
        model_configuration=ModelConfig(
            model_name="gpt-4o-mini",
            temperature=0.7,
            max_tokens=1000,
        ),
    ),
    fi_api_key=os.environ["FI_API_KEY"],
    fi_secret_key=os.environ["FI_SECRET_KEY"],
)

# Create the prompt as a draft and commit it as v1
prompt_client.create()
prompt_client.commit_current_version(
    message="Initial support prompt",
    label="production",
)

print(f"Created: {prompt_client.template.name} ({prompt_client.template.version})")

You should see:

Created: support-response (v1)

Check the dashboard: Prompts (left sidebar) → open support-response → click History → the History drawer lists v1 with the production label.

import os
import litellm
from fi.prompt import Prompt


def answer_question(question: str) -> str:
    prompt = Prompt.get_template_by_name(
        name="support-response",
        label="production",
        fi_api_key=os.environ["FI_API_KEY"],
        fi_secret_key=os.environ["FI_SECRET_KEY"],
    )

    # compile() returns a list of message dicts; pass to any LLM
    messages = prompt.compile(question=question)

    response = litellm.completion(
        model="gpt-4o-mini",  # swap for any litellm-supported model
        messages=messages,
    )
    return response.choices[0].message.content


print(answer_question("What is your return policy?"))

You should see a support-style answer printed to the terminal. compile() returns standard [{"role": "system", "content": "..."}, ...] message dicts, so it works with any litellm-supported model: swap in "groq/llama-3.3-70b-versatile" or "anthropic/claude-sonnet-4-20250514" and keep the rest of the function unchanged.

Each version can have its own model configuration. This v2 uses a lower temperature for more deterministic chain-of-thought responses.

from fi.prompt.types import PromptTemplate, SystemMessage, UserMessage, ModelConfig

# Create a new version with updated messages and model config
prompt_client.create_new_version(
    template=PromptTemplate(
        name="support-response",
        messages=[
            SystemMessage(
                content="You are a precise customer support agent for TechStore.\n\n"
                        "Think through the customer's question step by step before answering:\n"
                        "1. What is the customer asking?\n"
                        "2. What information do I have that directly addresses this?\n"
                        "3. What is the clearest, most helpful response?"
            ),
            UserMessage(
                content="Customer question: {{question}}\n\nAnswer:"
            ),
        ],
        model_configuration=ModelConfig(
            model_name="gpt-4o-mini",
            temperature=0.3,
            max_tokens=1000,
        ),
    ),
    commit_message="Add chain-of-thought reasoning",
)

# Save and commit v2
prompt_client.save_current_draft()
prompt_client.commit_current_version(message="v2: chain-of-thought prompt")

print(f"v2 created: {prompt_client.template.version}")

You should see:

v2 created: v2

v2 exists as a committed version but carries no label yet, so Step 2’s get_template_by_name(label="production") still serves v1.

Use is_concise here: for a support agent, concise answers are a key quality signal. Swap in any of the 72+ built-in eval metrics like groundedness, tone, completeness, or instruction_adherence depending on what you want to measure. Score v1 alongside v2 so you have something to compare before promoting.

import litellm
from fi.evals import evaluate

test_cases = [
    "What is your return policy?",
    "How long does standard shipping take?",
    "Can I exchange a product instead of returning it?",
]

v1_prompt = Prompt.get_template_by_name(
    name="support-response",
    version="v1",
    fi_api_key=os.environ["FI_API_KEY"],
    fi_secret_key=os.environ["FI_SECRET_KEY"],
)
v2_prompt = Prompt.get_template_by_name(
    name="support-response",
    version="v2",
    fi_api_key=os.environ["FI_API_KEY"],
    fi_secret_key=os.environ["FI_SECRET_KEY"],
)

print(f"{'Question':<45} {'v1':>8} {'v2':>8}")
print("-" * 63)

v1_pass = v2_pass = 0
for question in test_cases:
    row = [question[:43]]
    for label, prompt in (("v1", v1_prompt), ("v2", v2_prompt)):
        messages = prompt.compile(question=question)
        response = litellm.completion(model="gpt-4o-mini", messages=messages)
        output = response.choices[0].message.content

        result = evaluate("is_concise", output=output, model="turing_small")
        row.append(str(result.passed))
        if label == "v1" and result.passed:
            v1_pass += 1
        if label == "v2" and result.passed:
            v2_pass += 1
    print(f"{row[0]:<45} {row[1]:>8} {row[2]:>8}")

print(f"\nv1: {v1_pass}/3 concise   v2: {v2_pass}/3 concise")

You should see (illustrative, your model’s phrasing will vary):

Question                                            v1       v2
---------------------------------------------------------------
What is your return policy?                       True     True
How long does standard shipping take?             True     True
Can I exchange a product instead of returni      False     True

v1: 2/3 concise   v2: 3/3 concise

result.score is a float in [0, 1]. result.passed is the boolean pass/fail derived from it (score >= 0.5). Here v1’s answer to the exchange question ran long and failed is_concise, while v2’s step-by-step instruction kept all three answers concise. That 2/3 → 3/3 delta is the evidence for promoting v2, not just “it should work better.”

Prompt.assign_label_to_template_version(
    template_name="support-response",
    version="v2",
    label="production",
    fi_api_key=os.environ["FI_API_KEY"],
    fi_secret_key=os.environ["FI_SECRET_KEY"],
)

print("v2 is now live in production.")

You should see v2 is now live in production.

Your application now serves v2 on the next request, no redeploy. The get_template_by_name(label="production") call in Step 2 automatically picks up the new version.

If v2 causes issues, reassign the production label back to v1. Your app picks up the change on the next request.

Prompt.assign_label_to_template_version(
    template_name="support-response",
    version="v1",
    label="production",
    fi_api_key=os.environ["FI_API_KEY"],
    fi_secret_key=os.environ["FI_SECRET_KEY"],
)

print("Rolled back to v1.")

You should see Rolled back to v1., and the next call to answer_question() serves v1 again.

versions = prompt_client.list_template_versions()

for v in versions:
    draft = "draft" if v.get("isDraft") else "committed"
    print(f"  {v['templateVersion']}  {draft}  {v['createdAt']}")

You should see both versions listed (illustrative: list_template_versions() returns the backend history verbatim, so order isn’t guaranteed and your timestamps will match when you ran this tutorial):

  v2  committed  2026-08-18T09:15:00Z
  v1  committed  2026-08-18T09:10:00Z