Why this is in the vault
A plain-language survey of what "ontology" now means in the agentic data stack (machine-readable domain concepts, relationships and rules) and how the big vendors and two open-source repos implement it, ending on the open question of whether it beats a good agent plus a semantic layer.
The core argument
- Ontology is old (philosophy, 1960s CS, 1990s DARPA); it is resurfacing because agents need business context. Working definition: a machine-readable model of a domain's concepts, relationships and, where needed, rules.
- Rough split offered: a semantic layer answers "how do you calculate it"; an ontology answers "what is this thing". The author doubts the distinction lasts and expects the two to merge, with marketing confusion in the meantime.
- It is the latest in a long line of attempts to codify business knowledge (ERDs, data dictionaries, knowledge graphs, catalogs, semantic layers, data contracts). Possible catalog-style "implementation war" ahead.
- Open source is thin. NVIDIA auto-ontology crawls database metadata, has an LLM draft terms, SQL attributes and join relationships, humans certify, answers come back with their SQL. Judged heavy, SQL-only, no multimodal (PDF/Docs) ingestion, and yet another service to run.
- AWS context-ontology-accelerator is deeply AWS-bound (DataZone, Neptune, Bedrock and others) and costly up front, but builds graphs from PDFs and Word files.
- Databricks Genie Ontology (GA) is framed as a "context layer": human-defined knowledge plus auto-extracted snippets, ranked by authority, relevance, certification and recency. Same crawl, discover, human-refine pattern as NVIDIA's.
- Author's real question, left unanswered: is an agent with a heavy ontology better than a good agent with a semantic layer? He wants evals showing the lift justifies the infra.
- Conclusion: ontology is the old "what does customer mean" problem, now faced by an LLM instead of a human.
Mapping against Ray Data Co
Medium. The sharpest connection is the author's unanswered eval question: it matches RDCO's own typed knowledge graph (graph.duckdb, built from vault frontmatter and wikilinks) and the vault-note schema, a lightweight, human-curated ontology-in-practice for agents with no extra infra. It supports the stance that a thin, certified layer of entities and relationships beats a heavy platform, and it is a ready contrast for phData-facing conversations where Genie Ontology vs Snowflake semantic views will come up (the author calls Snowflake's approach "weird" without detail). Nothing here changes a project direction; it is vocabulary and a vendor-positioning snapshot. No deep-fetches (thought-leadership, not curation). Vendor claims are summarized from the author, not verified against vendor docs.
Related
- [[2026-04-04-ontology-taxonomy-knowledge-graphs]]
- [[2026-08-14-data-engineering-weekly-agent-ontology]]
- [[2026-08-26-data-engineering-weekly-operational-ontology]]
- [[2026-06-24-data-engineering-central-semantic-layers-agents]]