Amy Hodler, founder and executive director of GraphGeeks and co-author of O'Reilly's Graph Algorithms, argues that graphs matter now because they capture what sequence-predicting AI misses: the structure of relationships, and eventually the causes behind events. She ends forecasting graphs as agent memory within six months and causal graphs as the next big topic in about two years.
What a graph is. Hodler describes a graph as entities (nouns) joined by relationships (verbs), the circles and lines anyone draws on a whiteboard. Nodes are also called vertices, and edges are also called links. Property graphs attach properties to nodes and relationships: descriptors, quantities, normalised 0–1 strengths that summarise repeated interactions, lat-long codes, even lists. RDF triples (subject, predicate, object) come from the semantic web. They are more verbose and enforce more rules and logic, while property graphs are more flexible, which can help a team or get it into trouble.
Use cases. Graphs suit crime detection: money laundering, fraud rings, international criminal organisations and supply-chain crime. Bad actors imitate normal behaviour, and their pattern only shows when touch points are traced over time and across collaborators. Recommendations draw on shared directors and on the behaviour of similar viewers. Supply chains, she argues, failed in COVID because predictions rested on past correlations, while graphs find next-best routes when a port or warehouse goes down.
Algorithms and cost. Hodler calls graph algorithms her favourite part of the field because they compute over topology. They show whether a network is growing, clumping or dying, which statistical and ML methods miss. PageRank now appears in telomere and contaminant-toxicity research. On cost, she blames naive queries. All-pairs shortest path is "computationally crazy"; restricting it to the top k helps, as does tailoring queries and the data model to common questions. Distance can mean hops, time or cost, and in a relational database each hop is a join.
Where the money is. She calls fraud evergreen but crowded, and says cybersecurity is hot again amid an exponential rise in cybercrime. The low-hanging fruit is documentation analysis that combines an LLM with a graph. In legal work, chains of clauses and cross-references can reveal obligations impossible to fulfil, and help desks have similar needs. Caterpillar used graphs for support documentation long before RAG.
GraphRAG. Hodler says vector RAG has plateaued because of a meaning gap between questions and answers, approximate matches, missing topology and weak multi-step reasoning. Graphs add structure, sequences and aggregation across sources. She would rather call the result "hybrid RAG": the best results she has seen run vector and graph search in parallel and blend them. On tools she stays vendor-neutral. A dedicated graph database is no longer required, since graph engines or layers can project graphs over table data.
What's next. She sees frameworks diversifying, with hyperscalers re-entering the field. Multimodal graphs model relationships within images and audio, for example ships converging or a siren's Doppler shift. Graph memory for agents, extending short context windows, is the area she expects to be "super hot by the end of the year". Causal graphs, whose directed links could explain why sequences occur, are a little further off.