ptuner 0.4.0

Python client library for the ptuner prompt-tuning API

ptuner

Python client for the ptuner prompt-tuning API.

Evaluate, compare and iterate on LLM prompts with dataset-driven benchmarks, exact-match scoring, and LLM-as-judge evaluation.

Hosted at prompts.church

Installation

pip install ptuner

Quick Start

from ptuner import PtunerClient

client = PtunerClient(
    base_url="https://api.prompts.church",
    api_key="sk_...",
)

# 1. Create a project
project = client.create_project(
    name="Sentiment Analysis",
    description="Classify customer feedback",
)

# 2. Create a prompt with a version
prompt = client.create_prompt(
    project["id"],
    name="Sentiment Classifier",
    slug="sentiment-v1",
)

version = client.create_version(
    prompt["id"],
    description="Basic one-word classifier",
    system_template=(
        "You are a sentiment classifier. "
        "Respond with exactly one word: positive, negative, or neutral."
    ),
    message_template="Text: {text}\n\nSentiment:",
)

# 3. Create a dataset
dataset = client.create_dataset(project["id"], name="Customer Reviews")

reviews = [
    {"text": "This product is amazing!", "label": "positive"},
    {"text": "Terrible quality, broke after one day.", "label": "negative"},
    {"text": "The package arrived on time.", "label": "neutral"},
]

for r in reviews:
    client.create_datapoint(
        dataset["id"],
        description=f"{r['label'].title()} review",
        message_params=[{"role": "user", "params": {"text": r["text"]}}],
        exact_match_label=r["label"],
        external_id=r["text"][:64],  # optional: dedup on re-import
    )

# 4. Store your LLM API key (one-time)
client.create_credential(
    provider="openai",
    api_key="sk-your-openai-key",
    display_label="My Key",
)

# 5. Run evaluation
run = client.create_eval_run(
    project_id=project["id"],
    prompt_version_id=version["id"],
    dataset_id=dataset["id"],
    model_config={"model": "gpt-5-nano", "provider": "openai", "temperature": 0.0},
    judge_config={"judge_model": "gpt-5-mini"},
    iterations=3,
)

# 6. Wait and check results
import time
for _ in range(30):
    status = client.get_eval_run(run["id"])
    if status["status"] in ("completed", "failed"):
        break
    time.sleep(2)

results = client.list_eval_results(run["id"])
exact = [r["exact_match_score"] for r in results if r.get("exact_match_score") is not None]
judge = [r["judge_score"] for r in results if r.get("judge_score") is not None]

if exact:
    print(f"Exact match accuracy: {sum(exact)/len(exact):.1%}")
if judge:
    print(f"Judge avg score: {sum(judge)/len(judge):.2f}")

Authentication

Pass either an API key or a Firebase JWT token:

# API key (recommended)
client = PtunerClient(base_url="https://api.prompts.church", api_key="sk_...")

# Firebase JWT
client = PtunerClient(base_url="https://api.prompts.church", token="eyJ...")

Generate an API key in the UI at Settings → Generate API Key.

Structured JSON Output

Force models to return structured JSON by adding json_schema when creating a prompt version:

version = client.create_version(
    prompt["id"],
    description="JSON sentiment + confidence",
    system_template="You are a sentiment expert. Return JSON with sentiment and confidence.",
    message_template="Text: {text}",
    json_schema={
        "type": "object",
        "properties": {
            "sentiment": {"type": "string", "enum": ["positive", "negative", "neutral"]},
            "confidence": {"type": "number"},
        },
        "required": ["sentiment", "confidence"],
        "additionalProperties": False,
    },
)

This works across all providers (OpenAI, Anthropic, Google) — ptuner translates the schema to each provider's structured output format automatically.

Omit json_schema (or set it to None) for plain text mode.

Comparing Prompt Versions

A common workflow: iterate on a prompt and compare versions against the same dataset.

v2 = client.create_version(
    prompt["id"],
    description="Expert phrasing variant",
    system_template="You are a sentiment analysis expert. Respond: positive, negative, or neutral.",
    message_template="Text: {text}\n\nSentiment:",
)

run_v2 = client.create_eval_run(
    project_id=project["id"],
    prompt_version_id=v2["id"],
    dataset_id=dataset["id"],
    model_config={"model": "gpt-5-nano", "provider": "openai", "temperature": 0.0},
    iterations=3,
)
# Compare results between v1 and v2 in the UI or via the API

API Reference

Client

Method	Description
PtunerClient(base_url, api_key=, token=, timeout=)	Create a client
client.close()	Close the HTTP connection

Supports context manager: with PtunerClient(...) as client:

User

Method	Description
get_me()	Get current user info
generate_api_key()	Generate a new API key

Projects

Method	Description
list_projects()	List all projects
create_project(name, description="")	Create a project
get_project(project_id)	Get project details
list_members(project_id)	List project members
add_member(project_id, email, role="editor")	Add a member

Prompts & Versions

Method	Description
list_prompts(project_id)	List prompts in a project
create_prompt(project_id, name, slug)	Create a prompt
list_versions(prompt_id)	List versions of a prompt
create_version(prompt_id, description=, system_template=, message_template=, json_schema=)	Create a version

Datasets & Datapoints

Method	Description
list_datasets(project_id)	List datasets
create_dataset(project_id, name)	Create a dataset
list_datapoints(dataset_id)	List datapoints
create_datapoint(dataset_id, description=, system_params=, message_params=, exact_match_label=, acceptance_criteria=, labels=, external_id=)	Add a datapoint (upserts if external_id already exists in the dataset)
update_datapoint(datapoint_id, **fields)	Update a datapoint
delete_datapoint(datapoint_id)	Delete a datapoint

LLM Credentials

Method	Description
list_credentials()	List stored credentials
create_credential(provider, api_key, project_id=, display_label=)	Store a credential
update_credential(credential_id, **fields)	Update a credential
delete_credential(credential_id)	Delete a credential
resolve_credential(project_id, provider)	Resolve which credential will be used

Eval Runs

Method	Description
create_eval_run(project_id, prompt_version_id, dataset_id, model_config=, judge_config=, iterations=1)	Start an eval run
get_eval_run(run_id)	Get run status
list_eval_results(run_id)	Get run results
list_project_runs(project_id)	List all runs in a project

Examples

See examples/benchmark_sentiment.py for a full end-to-end benchmark that compares multiple models with both plain text and structured JSON output.

License

MIT