promptum

Test LLMs Like a Pro.

Stop writing boilerplate to test LLMs. Start getting results.

---

What's This?

You're choosing between GPT-4, Claude, and Gemini for your product. You need to know which one actually handles your prompts better — not some generic benchmark, but your real tasks, your edge cases, your expected formats.

promptum turns that into a few lines of Python. One client for 100+ models, automatic retries, latency/cost/token tracking, structured validation — all async, all typed, zero config files.

session = Session(provider=client, name="my_test")
session.add_test(Prompt(
    name="basic_math",
    prompt="What is 2+2?",
    model="openai/gpt-3.5-turbo",
    validator=Contains("4"),
))
report = await session.run()
summary = report.get_summary()

No YAML. No config files. Just Python you can read in 30 seconds.

---

Quick Start

pip install promptum  # (or: uv pip install promptum)
export OPENROUTER_API_KEY="your-key"

import asyncio
from promptum import Session, Prompt, OpenRouterClient, Contains

async def main():
    async with OpenRouterClient(api_key="your-key") as client:
        session = Session(provider=client, name="quick_test")

        session.add_test(Prompt(
            name="basic_math",
            prompt="What is 15 * 7? Reply with just the number.",
            model="openai/gpt-3.5-turbo",
            validator=Contains("105"),
        ))

        report = await session.run()
        summary = report.get_summary()

        print(f"Passed: {summary.passed}/{summary.total}")
        print(f"Avg latency: {summary.avg_latency_ms:.0f}ms")
        print(f"Total cost: ${summary.total_cost_usd:.6f}")

asyncio.run(main())

---

Why promptum?

Most LLM testing is ad-hoc scripts that grow into unmaintainable messes. You end up with separate API clients per provider, hand-rolled retry logic, manual latency tracking, and validation scattered across files.

promptum replaces all of that with a single coherent API:

• 100+ Models via OpenRouter — one client for OpenAI, Anthropic, Google, and more
• Smart Validation — ExactMatch, Contains, Regex, JsonSchema, or write your own
• Automatic Retries — exponential/fixed-delay backoff with configurable attempts
• Metrics Tracking — latency, tokens, cost — automatically captured
• Async by Default — run tests in parallel with concurrency control
• Extensible — implement LLMProvider or Validator to plug in any model or validation logic
• Type Safe — full type hints, catches errors before runtime

---

Documentation

• Session & Testing — Session, Prompt, Report, Summary, TestResult
• Providers — LLMProvider, OpenRouterClient, Metrics, Retry, Exceptions
• Validation — Validator, ExactMatch, Contains, Regex, JsonSchema

---

Requirements

• Python 3.10+
• An OpenRouter API key (or implement your own provider)

---

Contributing

Found a bug? Want a feature? PRs welcome!

git clone https://github.com/deyna256/promptum.git
cd promptum
just sync     # Install dependencies
just test     # Run tests
just style    # Check code style
just format   # Format code
just type     # Type checking

---

License

MIT - do whatever you want with it.

---

Star on GitHub | Report Bug | Request Feature

Made for developers who value their time.