// FAQ
Questions, answered.
Everything about how HallucinationLab measures reliability — and what it does not do.
How do you detect hallucinations without another LLM?
Deterministically. Every answer is scored against the ground truth or source material you provide, using fixed scoring logic — not a second model's opinion. We never use one LLM to grade another, because a grader that can hallucinate cannot certify reliability.
Do you store my API keys?
No. Your provider keys live only in your browser session and are sent directly to the model providers from your machine. They never touch our servers and are never logged.
Which models can I compare?
Models across Anthropic (Claude Opus, Sonnet, Haiku), OpenAI (GPT-5.5, GPT-5.4 Mini), Google (Gemini 3.1 Pro, 3.5 Flash, Flash Lite), xAI (Grok 4.3), Perplexity (Sonar Pro), and Groq-hosted open models — side by side on the same prompt.
What are the four reliability metrics?
Faithfulness (is the answer grounded in the source?), accuracy (does it match the ground truth?), relevance (does it answer the question?), and hallucination rate (how often does it fabricate?). Together they describe how much you can trust a model on your data.
Is it free?
The evaluation tool is free to use — you bring your own API keys. The hlabs SDK is in private beta; join the waitlist for access, which is free for the duration of the beta.
What is the hlabs SDK?
A drop-in layer that sits between your application and any LLM. It scores each response for faithfulness in real time and automatically retries or falls back to a more reliable model when an answer does not hold up — so a wrong answer never reaches your user.
Do you support retrieval-augmented generation (RAG)?
Yes. When your app supplies source documents, faithfulness is measured against that context — exactly the setup most RAG applications run in production.
Still curious?
Try an evaluation →