Structured agentic retrieval over knowledge graphs with MCP and query containers

Integrating Large Language Models (LLMs) with Knowledge Graphs (KGs) remains challenging due to the difficulty of translating natural language questions into valid and semantically grounded graph-based queries. Free-form SPARQL generation is particularly prone to hallucinations and errors when operating over complex and domain-specific schemas. In this paper, we propose a framework for structured agentic retrieval over KGs based on the Model Context Protocol (MCP). Our approach replaces unconstrained generation with a tool-driven workflow in which queries are incrementally constructed through low-granularity operations. Central to this design is the Query Container, a stateful abstraction that maintains a consistent query state and enforces structural validity throughout the process. We further introduce a hybrid decision mechanism combining constrained tool usage with model-guided selection via micro-sampling, coupled with execution-based validation. Experimental results show improved robustness and a reduction in hallucination-related errors compared to standard generation approaches.