The invention relates to the technical field of
large model service tracking, and discloses a
large model service
chain type calling
tracking system and a
root cause analysis method, and the method comprises the steps: carrying out the semantic coding of the input and output of a
large model, generating a final
fingerprint, and carrying out the cross-Span
fingerprint inheritance and extension; semantic uncertainty is calculated and analyzed, and the sampling rate is dynamically adjusted; based on sampling data acquired by a sampling rate, fusing a call chain topological structure and a time
decay factor, constructing a dynamic space-time atlas, optimizing an atlas index, constructing a multi-
dimensional analysis framework, and realizing positioning from surface anomaly to deep
root cause in combination with an anomaly propagation contribution degree obtained by calculation; carrying out statistics on distribution parameters of local indexes and carrying out
differential privacy processing, and calculating a global threshold value of the distribution parameters and carrying out issuing execution; and based on a final positioning
root cause obtained by global threshold analysis, evaluating a root cause repair effect and performing model iteration, and converting historical root cause data into a driving force for
system optimization.