SearchAtlas Framework for Analyzing Agentic Search Trajectories
September 11, 2026
SearchAtlas converts LLM search trajectories into structured evidential query graphs to evaluate reasoning processes beyond final-answer accuracy. The automated pipeline achieves an 86.0% mean edge F1 score, exposing how agents aggregate evidence or fail to address constraints.
HOW THIS AFFECTS YOU
●
builderYou can use this to debug why your search agents fail to ground answers in retrieved evidence.
●
researcherThis offers a structured way to move evaluation from outcome-based to process-based metrics.