AI Systems / POV
Do We Really Need Knowledge Graphs in the Age of LLMs?
For most problems, I think knowledge graphs are over-engineering. A case for multi-agent systems with chain-of-thought reasoning as the more adaptable alternative.
My view is that for most problem statements, a knowledge graph is over-engineering. They are often proposed to reduce hallucinations, but building and maintaining one takes work: new data needs to be structured, connected and kept up to date. I describe my own attempt to build one for a recommendation engine, where the visualisation gave no real benefit and the time complexity did not change.
The alternative I suggest is a multi-agent system with chain-of-thought reasoning. Break the query into structured steps, let agents run RAG in parallel across sources, compare what comes back and use it to answer the question. In the use cases I worked on, I found this could give comparable results with less to maintain. I do not say graphs are never useful, only that they fit specific use cases and are not the default.
Summary of an essay originally published on Medium. Read the original ↗