CLAIM: Causal LLM-Assisted Inference for Microservices

Accurate causal graphs are critical for root cause analysis in microservice systems, yet purely data-driven methods struggle to distinguish direct fault-propagation edges from transitive correlations induced by service call chains. Scorebased causal discovery algorithms such as the Greedy Equivalence Search (GES) can improve accuracy when augmented with domain knowledge encoded as forbidden-edge lists, but authoring such constraints requires algorithmic expertise and system-specific engineering. We propose CLAIM (Causal LLMAssisted Inference for Microservices), a method that expresses the same domain knowledge as natural language instructions and combines it with pre-computed statistical evidence in a per-service Large Language Model (LLM) inference call. This formulation eliminates the need for algorithm-specific constraintimplementations while enabling operators to author and revise causal priors in plain text. On a synthetic benchmark with random Directed Acyclic Graph (DAG) topologies at three scales (4, 8 and 12 services), CLAIM achieves near-perfect graph recovery (F1 ≥ 0.939 with perfect recall) at all scales, exceeding the accuracy of GES with equivalent domain knowledge, while its runtime grows ×6 from 4 to 12 services where GES grows ×1873.