Agentic SPARQL: Evaluating SPARQL-MCP-powered Intelligent Agents on the Federated KGQA Benchmark
New framework combines LLM agents with SPARQL endpoints via MCP, automating complex data queries across multiple sources.
A research team led by Daniel Dobriy and Frederik Bauer has published a paper introducing 'Agentic SPARQL,' a novel framework that bridges the gap between large language model (LLM) agents and federated knowledge graphs. The system leverages the emerging Model Context Protocol (MCP), a standard for connecting LLMs to external tools, to empower AI agents to autonomously perform complex data retrieval. Specifically, it enables agents to discover available SPARQL endpoints, explore their metadata schemas, and formulate precise queries that can federate—or combine—data from multiple, disparate knowledge sources. This addresses a key challenge in knowledge graph question answering (KGQA), where relevant information is often siloed across different databases.
The researchers' work is both practical and evaluative. They first extended an existing Knowledge Graph Question Answering benchmark to create a new Federated KGQA (FKGQA) benchmark suitable for testing agentic systems. They then implemented and evaluated different architectural approaches for integrating SPARQL federation with LLM agents via MCP. This involved testing the agents' capabilities in the full pipeline: from selecting the right data sources and understanding their structure to ultimately generating and executing the correct federated SPARQL query. The paper positions this 'agentic' approach as a natural evolution and complement to prior work on automated query federation, harnessing LLMs' planning and reasoning abilities to navigate the complexity of interconnected data webs.
- Framework combines LLM agents with SPARQL endpoints using the Model Context Protocol (MCP) for standardized tool access.
- Enables autonomous agent tasks: endpoint discovery, schema exploration, and federated SPARQL query formulation across multiple knowledge graphs.
- Researchers created and tested against an extended Federated KGQA benchmark to evaluate different agentic architectures.
Why It Matters
Automates complex data retrieval from siloed sources, paving the way for AI agents that can answer intricate questions by synthesizing information from multiple knowledge graphs.