Snowflake and InfoQ will host a livestream on August 4 titled '2026 Data+AI Mid-Year Review' to discuss the reality of deploying enterprise agents. The article translates a blog by Snowflake architect Jia Tianxia, explaining how integrating ontologies into Cortex Agents enhances AI reasoning, with a biomedical benchmark demonstrating the benefits.
SynthePulse Insight · AI deep readingMembers
Version 1 · 1 source
When agents enter core enterprise processes, the semantic layer alone is insufficient for trustworthy reasoning. In a preview of a Snowflake and InfoQ livestream, a biomedical benchmark shows that ontology-aware approaches can significantly improve accuracy, but the real key may not be computational power—it's structured context.
Enterprise data is full of meanings not explicitly represented in database schemas. This implicit domain knowledge must be captured to support reliable agent systems. Many industries maintain formal ontologies, such as SNOMED CT and UMLS in healthcare, GS1 in supply chain, and FIBO in finance, published in formats like OWL, RDF, or SKOS, containing thousands of concepts and tens of thousands of hierarchical edges. However, most enterprise AI systems still operate on relational abstractions, unable to natively access knowledge layers like concept inheritance and transitive relationships, creating a gap between modeling real-world meaning and AI retrieval and reasoning.
Snowflake's Semantic View provides a foundation for relational constructs, but the semantic layer alone may not capture complete domain concepts like hierarchies and synonyms. To bridge this gap, Snowflake explores integrating ontologies into Cortex Agents, enabling agents to understand business concepts and their relationships.
This article is based on a translation of a blog post by Jia Tianxia, Head of Global AI Technology Strategy at Snowflake, published by InfoQ. The benchmark was designed internally by Snowflake, and results have not been independently verified by third parties. The author explicitly states the directional, controlled conditions. All numbers and conclusions come from this single source and should be treated as source claims rather than independent confirmations.
Primary report
Primary source