Snowflake’s ontology benchmark says the data layer—not more agent tools—drove most of the gain

Snowflake has published a useful controlled comparison of four Cortex Agent architectures for questions that cannot be answered from tables and joins alone. On a 22-question biomedical test, its standard semantic-view baseline scored 50% accuracy. Adding a knowledge graph raised that to 60%; a flattened GraphRAG index reached 70%; and GraphRAG plus eight curated terminology mappings reached 78.2%. Each configuration was run five times. Snowflake’s engineering post describes the result as directional evidence, not a production benchmark.

The simpler agent beat the richer toolset

The comparison matters because the knowledge-graph agent had seven tools: four stored procedures, two specialized Cortex Analyst semantic views and the baseline machinery. The flattened GraphRAG agent had only two—a Cortex Search retrieval tool and a Cortex Analyst SQL tool. Snowflake says the latter produced higher average accuracy and the lowest observed run-to-run variance.

The difference was largely moved upstream. Rather than asking the agent to traverse an ontology at query time, Snowflake precomputed a denormalized profile for each concept. Those profiles combined names, definitions, synonyms, nearby graph structure and attributes aggregated from descendants, then indexed them for hybrid keyword and vector retrieval. A retrieved profile supplied explicit categories to the SQL step.

That is a practical design choice for analytics teams. More graph tools can provide more raw capability, but they also give the orchestrator more opportunities to choose the wrong operation or sequence. Precomputing the relevant hierarchy into searchable context narrows the agent’s decision space before SQL generation.

The last gain came from hand-built mappings

The strongest configuration added eight authoritative cell-type-to-tissue mappings and definitions for composite terms to the system prompt. That lifted measured accuracy from 70% to 78.2%, but Snowflake also says this version had the highest variance. The mappings helped with phrases whose intended categories could not be recovered from hierarchy traversal alone.

That makes the top score less general than a headline number might suggest. It combines an architectural improvement with task-specific curation. Snowflake’s own recommendation is appropriately restrained: keep that hardcoded layer thin and use it to supplement a complete index, not replace one.

What the benchmark does—and does not—show

The test used a simplified biomedical dataset joining the Cell Ontology—33,651 terms and roughly 50,000 edges—with PRISM drug-screening data covering 4,518 drugs, 578 cancer cell lines and more than 2.6 million viability measurements. The 22 hard questions were deliberately designed around ontology-dependent reasoning. That is enough to compare these four implementations under the stated conditions, but not enough to establish general Cortex Agent accuracy.

The defensible takeaway is narrower: when business questions require taxonomy expansion, aliases or inherited categories, a semantic layer may need a second structural grounding layer. Before adding more runtime tools, teams should test whether precomputed concept profiles can make the required meaning explicit. Snowflake’s result suggests that fixing retrieval context can deliver most of the observed improvement while simplifying orchestration.