Does groundedness check every LLM call, or just the final answer?

Every LLM call. It scores each call against the context that call received as the call happens, not just the last message a trace sends back to the user.

Multi-step agents make several calls before they ever reply: planning a next action, reasoning about which tool to use, summarizing a tool result. An agent can invent a detail two steps before the final answer, and if that detail feeds what generates the reply, the final message can read as perfectly confident and cite nothing that looks wrong on its face.

Checking only the last message would miss that: the final answer can be a faithful, well-written summary of an earlier hallucination. Scoring every call catches the step where the claim was actually invented, with that specific call attached to the finding, instead of only noticing something’s off once it’s already baked into what the user sees.

keep reading

More on this.

Can a factually correct answer still get flagged as ungrounded?Yes. Groundedness checks whether an answer is supported by its context, not whether the answer is true, and those are different questions.answer →Does the groundedness classifier check whether the context itself is correct?No. It checks whether the answer matches the context given, not whether that context is accurate, so a wrong or stale source can still score grounded.answer →Does the groundedness classifier only work on agents that use RAG?No. It scores any LLM call against whatever context that call received, tool outputs and prior turns included, not only retrieved documents.answer →Everything on tessary groundednessThe definition, and every question answered under tessary groundedness.concept →

Send us the traces you already emit.

book a demo