Veritas · LLM gateway

A semantic cache you can prove is correct.

A provider-agnostic LLM gateway whose two-tier cache is measured the way the commercial tier never is: not just a hit rate, but the precision and false-positive rate of those hits — so a paraphrase reuses an answer, but a near-identical question with the opposite answer does not.

Most semantic caches report only how often they hit. GPTCache, for one, publishes hit-ratio and recall but no precision — so a high hit rate can quietly include wrong answers. Veritas scores the metric they omit and gates CI on it.

· Two-tier cache: exact (scoped hash) + semantic (embedding cosine)
· An eval harness scores cache-hit precision, recall, and false-positive rate
· Adversarial golden set: paraphrases should hit; negations must not
· Provider-agnostic over Anthropic + OpenAI, with a deterministic offline path

Open the live playground →

Runs fully offline with no API key (deterministic mock provider + committed eval embeddings). Add an ANTHROPIC_API_KEY or OPENAI_API_KEY to route to a live model.