variA/Bly is a deterministic AI evaluation platform. It scores LLM and RAG outputs across 49 dimensions in 6 categories — claim-level grounding, hallucination detection, retrieval quality, coherence, and more — producing the same score every run.

variA/Bly has no LLM judge in the grading path, so scoring is deterministic and reproducible. On the public RGB benchmark it wrongly approves ungrounded answers 6.1% of the time, versus 38.2% for RAGAS.

Yes. variA/Bly uses Natural Language Inference, embedding similarity, and rule-based checks instead of a judge model, so the same input always produces the same score — which matters for regression testing and audits.

RAG systems, single LLM outputs, and multi-agent workflows — including per-hop grounding across agent handoffs, domain-aware routing for clinical, legal, and finance text, and multilingual output.

A subscription plus pay-as-you-go per-evaluation (SEU) pricing, with a free tier to start. It is typically 50–60% less than the LLM-judge cost of RAGAS alone.

Any LLM behind an HTTP API — OpenAI, Anthropic, Google, Meta, Mistral, and your own self-hosted models. variA/Bly is provider-agnostic: you call your model however you like, then score the response with one SDK call.

Through a lightweight SDK that adds under 1 millisecond to your request. Scoring runs asynchronously on variA/Bly’s hosted infrastructure, so it never blocks your response, and setup takes about 10 minutes.

Yes. It scores every agent hop and the handoffs between them, so when a multi-agent workflow produces a wrong answer you can see exactly which step introduced the drift — not just that the final output was wrong.

Claim-level grounding breaks a response into individual factual claims and checks each one against the retrieved source documents, flagging any claim that is not supported. It is how variA/Bly catches hallucinations that a whole-response score would miss.

Deterministically — using Natural Language Inference for entailment, numeric verification, and contradiction checks against the source, rather than asking another LLM for an opinion. Every flagged claim comes with the source reference and the reason it failed.

An SEU is variA/Bly’s usage measure. Most evaluation runs consume 1–3 SEUs, which keeps billing predictable regardless of prompt size or judge-model price changes.

Observability tells you what your LLM did — traces, latency, logs. variA/Bly tells you whether the answer was actually grounded and correct, deterministically. Many teams run both: observability for what happened, variA/Bly for how good it was.

Yes. The per-sample scores from the RGB benchmark are committed to the public benchmark repo, so you can recompute the headline numbers yourself with no account and no cost, or re-run the scorers from scratch.

Sign up for a free API key and score your first evaluation with the SDK in a few lines, or email info@variably.dev and we will help you get going.