Descriptions:
Tim Ainge of The Good Collective delivers a practitioner-focused guide to graph data structures for AI application developers at the AI Engineer conference. Rather than covering GraphRAG or agent memory graphs, the talk focuses on foundational concepts — nodes, edges, schemas, ontologies — and how they combine with modern AI techniques to power search, retrieval, and recommendation use cases that would be unwieldy in relational databases.
The talk walks through a concrete progression using recipe extraction as a running example: starting with naive triple extraction (subject, predicate, object), then improving results dramatically by providing the extractor a typed schema to fill, then adding ontology rules for standardization (lowercase ingredient names, metric units), and finally using embedding models to solve the entity-matching problem — identifying that “potato” and “potatoes” are the same node — without needing a fixed vocabulary defined in advance. Code examples accompany each stage, and Ainge explicitly frames the hybrid of graph structure plus embedding-based matching as producing better results than either technique alone.
The second half covers graph-native algorithms, centering on personalized PageRank (PPR): a variant of Google’s original PageRank that teleports back to a starting node to surface contextually relevant neighbors. Ainge references the Pinterest Pixie paper as a canonical application and HippoRAG as a recent AI-era use case linking memories to questions. The talk positions graphs not as a universal solution but as a powerful complement to AI in problems involving multi-hop relationships and structured knowledge retrieval.
📺 Source: AI Engineer · Published July 18, 2026
🏷️ Format: Deep Dive
