Learn · Demo
Evaluating LLM answers
"It sounds right" isn't good enough for enterprise AI. Pick a task, then compare candidate answers on two real signals: overlap with a reference answer (token-F1) and grounding in the source context — the tell-tale for hallucination.
Choose a task:
Source context
Question
Reference answer (ground truth)
Score a candidate answer:
Candidate answer
—Pick a candidate above.
■ supported by context · ■ not found in context (possible hallucination)
Token-overlap F1 vs reference
—Precision × recall of content words shared with the reference answer.
Grounding vs context
—Share of the answer's content words actually found in the source context.
Why two metrics?
F1 tells you whether the answer matches what a human said — but a fluent wrong answer can still score low. Grounding catches the dangerous failure mode: confident claims that aren't in the source at all. Low grounding is the signature of a hallucination. Real eval pipelines add semantic similarity and LLM-as-judge scoring on top, but the principle is the same — never trust fluency alone.
This demo uses word-overlap maths (content words, stopwords removed) so you can see exactly how the scores are computed. Production evaluation also uses embedding-based semantic similarity — see the Embeddings demo — and human or model graders for nuance.