Quality dimensions
The semantic axes HelixQA scores.
LLM-based judges score each dimension with attached reasons; deterministic checks run in parallel.
- Correctness — is the answer factually right for the given input?
- Relevance — does it address the user's actual request?
- Groundedness — is it supported by the provided context / retrieved documents?
- Instruction following — does it obey the system prompt and format rules?
- Safety — does it avoid harmful, off-policy or disallowed content?
- Regression detection — across runs, which tests got worse and by how much?
Dimension scores in the UI and API come only from backend responses — the frontend never invents scores.