Summary: Representing knowledge as entities and relationships (an ontology) rather than documents. Hard for humans to read, but valuable to agents for understanding how an organization works: who, what, when, why, and what depends on what. A graph gives agents a map instead of a flashlight, so they don’t exhaust themselves querying one spot at a time or misread an empty result from a slightly wrong filter as “nothing happened”. Popularized in enterprises by Palantir; GraphRAG and Databricks’ Genie Ontology bring it to AI stacks, and security platforms build hybrid graphs (batch-built, plus live queries for freshness).
Key ideas
- Ontologies help agents grasp an org’s relationships, metrics, and mechanics (ai-engineering-podcast—company-brain, roi-zalta).
- Palantir deploys engineers to build company-level ontologies for defense, logistics, and more (ai-engineering-podcast—company-brain).
- Databricks Genie Ontology ranks “snippets” (docs, tables, dashboards, markdown, people) with a relevance model (OntoRank) and can tell an agent “this only lives with this person; go ask him” (ai-engineering-podcast—company-brain).
- Key elements to extract: actors, interactions, time (relative urgency), rationales, dependencies (ai-engineering-podcast—company-brain).
- Maintenance options include keeping only metadata/references plus cleanup jobs (ai-engineering-podcast—company-brain).
- Topology maps help agents: a customer-validated map of Kubernetes resource relationships, serialized as YAML in the prompt, beat letting the agent rediscover them (langtalks—65-ai-sre).
- Humans organize knowledge into schemas that speed encoding and retrieval. Agent memory stores facts independently, and it’s unclear that GraphRAG matches how the brain does this (langtalks—60-brain-memory).
- Give agents a map, not a flashlight: query tools let an agent illuminate one spot at a time, and it exhausts itself or stops at the first hit. A context graph of resolved entities and relationships (an ontology over normalized data) answers in one place. Build it in batches, and combine it with live queries for freshness (hidden-layers—vega-gili-kanfo, gili-kanfo).
- Silent failures motivate it: a query filtering on a slightly wrong value (“uri” vs “Uri”) runs fine and returns nothing, which an agent reads as “no activity”, whereas a human would sense something off (hidden-layers—vega-gili-kanfo, gili-kanfo).
Disagreements & open questions
Takeaways
Related
second-brain · rag · ai-sre