FAQ¶
Frequently asked questions about OpenAgent Eval.
What is OpenAgent Eval?
An open-source, local-first framework for evaluating RAG systems and AI Agents. Our goal is to
become the pytest of AI evaluation — a familiar, composable way to measure quality.
Do I need an API key?
Only if you use a hosted LLM provider (OpenAI, Gemini, Anthropic, Groq, OpenRouter). You can run
fully locally with Ollama, or use the built-in mock providers for CI and
dry-runs — no API key required.
Which LLM providers are supported?
OpenAI, Google Gemini, Anthropic, Groq, OpenRouter, Ollama, and a mock provider. See
Architecture for the base classes and how to
add your own.
Which retrievers are supported?
Chroma, Memory, BM25, FAISS, Qdrant, Pinecone, Weaviate, Elasticsearch, PGVector, HTTP, and a
mock provider. Implement Retriever to add one.
Which embedders are supported?
Sentence-Transformers and a mock embedder. Local vector retrievers (Memory, FAISS, Qdrant,
Pinecone, PGVector) use an embedder; server-side backends (Chroma, Weaviate) embed remotely.
Can I use OpenAgent Eval inside pytest?
Yes. Construct an Engine and call its async run() method with asyncio.run. See
Examples.
How do I add a custom metric?
Subclass openagent_eval.metrics.base.BaseMetric, implement evaluate(**kwargs) -> MetricResult,
and register it in METRIC_REGISTRY. A template lives at
openagent_eval/plugins/examples/custom_metric.py.
Which report formats are available?
Terminal, Markdown, HTML, and JSON, plus a side-by-side comparison report for
oaeval compare.
Does it send my data anywhere?
No. OpenAgent Eval is local-first. The only network calls are to the LLM/retriever providers you configure. We do not collect telemetry.
What Python versions are supported?
Python >= 3.11.
How do I deploy the documentation?
Push to main — GitHub Actions builds and deploys the site to GitHub Pages automatically. See
Contributing for local preview.
Where do I get help?
- GitHub Issues for bugs
- GitHub Discussions for questions
- Discord for community chat