Ground the agent to reduce hallucination

Anchor the agent in real data and require it to cite sources, so it reports what is true instead of inventing what sounds plausible.

Give it the facts, do not rely on memory

A model asked to recall details from training will sometimes confabulate. An agent given the relevant documents and asked to answer only from them is far more reliable. Retrieval-based grounding is the standard tool here: fetch the source, then answer from it.

Require citations and verify them

Ask the agent to point at the specific source for each claim, then check that the source says what the agent claims. Faithfulness metrics from tools like Ragas can flag answers that are not supported by the retrieved material.

Let it say "I do not know"

A large share of hallucination happens because the agent feels obligated to answer. Explicitly permitting and rewarding "I do not have that information" removes the pressure that produces made-up answers.