A service-oriented architecture for feedback-driven reconfiguration of LLM analytics pipelines

Analytics pipelines increasingly compose heterogeneous services—retrieval modules, LLM endpoints, vector stores, validators, and delivery components—yet their service configurations remain static even as costs, quality, and governance requirements change at runtime. This paper proposes a service-oriented architecture in which a policy-bounded LLM planner, embedded within a MAPE-K control loop, reasons over telemetry signals, structured service descriptors, and task intent to recommend verified reconfiguration actions—including service substitution, traffic rerouting, context-budget tuning, validator insertion, and graceful degradation. We formalise the service graph model, define a typed operator catalogue with explicit pre- and post-conditions, and illustrate the design through a document-grounded analytics scenario. The contribution is a precise, service-centric design proposal for self-adaptive LLM pipelines, accompanied by an evaluation agenda covering operational, economic, quality, and governance dimensions.