Skip to main content
aayushprime.

Note

Agent Evals

notes from this short video.

We want agents to be reliable. Reliability = Observability + Evals

Observability means we can see what agent did. What the model response was, what the tool response was.

LLM response desired properties:

  • Relevance: Answer relevance to the user’s query

  • Groundedness (relates to hallucination as well): uses the particular context provided (as opposed to the generic response)

  • Tool Sequence Accuracy: does not do wasteful tool calls, and follows correct sequence of tool calls

Evals = QA for agents

Who does evals? Whoever builds the app, also runs evals.

make a synthetic dataset to run evals (in the beginning to kickstart the eval process) or a golden dataset (domain expert designed)