DeepEval v3: Component-Level LLM Evaluation with Tracing — Test Individual Retrievers, Tool Calls, and Agent-to-Agent Interactions, Not Just End-to-End

DeepEval·medium signal

DeepEval v3.0 adds component-level evaluation that assesses individual LLM workflow units — retrievers, tool calls, generators, agent interactions — instead of treating the app as a black box. Spans within a trace represent component-level evals. Combined with production observability and conversation simulation, this enables debugging at the layer where failures actually occur. Key pattern: build a golden dataset from real production failures (200-500 examples), not synthetic data.