What is Tessary's groundedness classifier?

It’s a built-in classifier that checks your agent’s answers against the documents it retrieved, or against the prompt when it retrieved none, and flags the sentences that source doesn’t support.

A sentence counts as unsupported when the source contradicts it or never says it at all. The scorer is an open model, tessaryai/groundedness-classifier-v1, which reads the source and the whole answer together in one pass and scores each sentence for how likely it is to be unsupported. It isn’t an LLM judge: it runs on a GPU you provide, a Mac for development or an AWS instance for production, with no provider key and no per-answer charge.

On the RAGTruth benchmark, its model card reports that at 0.975, the threshold Tessary flags at, it catches 39% of unsupported answers, and 83% of the answers it flags are unsupported. So one flag isn’t a finding. Each call site learns its normal share of traces with a flagged answer, and a finding opens when that share rises, then goes to triage.

The limit is recall: at that threshold it misses most unsupported answers, so a quiet call site doesn’t mean every answer on it was grounded.