AutoSchema Improves Text-to-SPARQL Querying for Diverse Knowledge Graphs

Traditional methods for AI agents to query complex knowledge graphs, especially in life sciences, often depend on meticulously crafted metadata files. These files, like TogoMCP, help language models understand the diverse schemas and identifiers across different data sources. However, creating and maintaining these metadata files is a labor-intensive process, requiring significant human and AI assistance for drafting, validation, and review. A new approach, called "live schema grounding," aims to overcome this limitation. This method allows an agent to directly obtain the necessary schema information from the live endpoints of the knowledge graphs as needed for a specific query. The proposed framework, AutoSchema, implements this by inspecting live schemas, mapping entity names from a question to graph identifiers, exploring relation paths, and dynamically finding connections between resources during iterative query construction. Evaluations against the TogoMCP framework show that AutoSchema significantly improves factoid accuracy in biomedical knowledge graph question answering and offers consistent gains in longitudinal BioASQ evaluations. It also reduces the number of tool calls and budget exhaustion, demonstrating greater efficiency. Preliminary studies further suggest its capability to handle irregular and previously undocumented graphs without requiring prior schema curation.