openeval-ai 0.1.0

   ____                   ______          __
  / __ \____  ___  ____  / ____/___  ____/ /____
 / / / / __ \/ _ \/ __ \/ __/ / __ \/ __  / ___/
/ /_/ / /_/ /  __/ / / / /___/ / / / /_/ (__  )
\____/ .___/\___/_/ /_/_____/_/ /_/\__,_/____/
    /_/

Model-agnostic LLM evaluation that feels like pytest.

Install

pip install openeval-ai

Optional extras:

pip install "openeval-ai[openai]"      # OpenAI adapter
pip install "openeval-ai[anthropic]"   # Anthropic adapter
pip install "openeval-ai[semantic]"    # semantic_sim scorer

Quickstart (10 lines, zero config)

from openeval import EvalCase, run
from openeval.scorers import exact_match

cases = [EvalCase(input="2+2", expected_output="4")]

def my_model(prompt: str) -> str:
    return "4"

summary = run(my_model, cases, {"exact": exact_match})
print(summary.pass_rate)  # 1.0

Why OpenEval

• Scorers are just Python functions — composable, testable, versionable.
• Model-agnostic — OpenAI, Anthropic, Ollama, or any callable.
• CI-first — pytest plugin + GitHub Action for evals on every PR.

Architecture

flowchart LR
  M[Your Model] --> A[Adapter]
  A --> R[Runner / Async Runner]
  R --> S[Scorers]
  S --> E[EvalSummary]
  E --> H[HTML Report]
  E --> C[CLI]
  E --> G[GitHub Action]

Comparison (benchmarks)

Benchmarks live in benchmarks/ and use a fully offline mock model for reproducibility.

Measured on 2026-03-20 with uv run benchmarks/run_all.py (see benchmarks/results.json).

Task A: 20-case Q&A (exact match), mock model with 10ms latency per call.

Notes:

• Promptfoo is optional in the harness; set OPENEVAL_BENCH_PROMPTFOO=1 to run it.
• Promptfoo may require Node.js >=20.20.0 or >=22.22.0 (newer than v22.17.x).

Built-in scorers

Built-in suites

Run a suite:

from openeval.suites import factuality
summary = factuality.run(lambda _: "Paris")
print(summary.pass_rate)

GitHub Action

# .github/workflows/eval.yml
name: LLM Eval
on: [pull_request]
jobs:
  eval:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: OWNER/openeval-action@v1 # replace OWNER
        with:
          suite: hallucination
          model: openai:gpt-4o-mini
          threshold: "0.8"
        env:
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

CLI

# Run a suite
openeval run --suite factuality --model ollama:llama3.2

# Parallel (async) + HTML report
openeval run --suite hallucination --model openai:gpt-4o-mini --async --concurrency 20 --output report.html

Async evaluation

import asyncio
from openeval import EvalCase, arun
from openeval.scorers import exact_match

cases = [EvalCase(input="2+2", expected_output="4")]

async def model(_prompt: str) -> str:
    return "4"

summary = asyncio.run(arun(model, cases, {"exact": exact_match}, concurrency=20))
print(summary.pass_rate)

Examples

• examples/01_basic_qa.py
• examples/02_openai_eval.py
• examples/03_compare_models.py
• examples/04_rag_eval.py
• examples/05_chatbot_eval.py
• examples/06_code_eval.py
• examples/07_custom_scorer.py
• examples/08_cicd_integration.py

Contributing

See CONTRIBUTING.md.

License

MIT — see LICENSE.