verdikt-sdk

Python SDK for Verdikt — a standalone AI evaluation service that decouples evaluation and LLM/human judging from the application being evaluated.

Installation

pip install verdikt-sdk

Usage

from verdikt_sdk import AnswerWithCost, VerdiktClient, EvaluationType, Question
from yalc import LLMModel

client = VerdiktClient(
    base_url="https://your-verdikt-instance.com",
    client_id="your-client-id",
    client_secret="your-client-secret",
)

# Register your app (idempotent — safe to call on every deploy)
await client.create_app(slug="my-app", name="My App")

# Sync questions to the dataset (idempotent)
await client.add_questions("my-app", [
    Question(question="What is the capital of France?", human_answer="Paris"),
])

# Your callback returns the answer plus the cost it took your app to produce it.
# `cost` is optional — pass None when you do not track it.
async def my_llm_function(question: str) -> AnswerWithCost:
    answer, cost = await my_app(question)
    return AnswerWithCost(answer=answer, cost=cost)

# Run an evaluation cycle
await client.run_evaluation(
    app_slug="my-app",
    app_version="v1.2.0",
    callback=my_llm_function,
    evaluation_type=EvaluationType.LLM_ONLY,
    llm_judge_models=[LLMModel.gpt_4o_mini],
)

run_evaluation calls your callback concurrently for every question in the dataset, then submits all answers to Verdikt for judgment.

Breaking change in 0.2.0: the callback now returns AnswerWithCost(answer=..., cost=...) instead of a bare str. Callers on 0.1.x must wrap their return value (return AnswerWithCost(answer=ans) is a drop-in equivalent of the old behaviour).

Authentication

The SDK authenticates via Zitadel OAuth2 client credentials. Create a machine user in your Zitadel project and pass its client_id and client_secret to EvaluationClient.