From Similarity to Structure: A Decision Framework for Graph-Aware Retrieval in Enterprise AI Systems

Retrieval-Augmented Generation (RAG) has become the default pattern for grounding large language models in enterprise knowledge, yet most production pipelines index that knowledge as flat text and retrieve it by vector similarity alone. This works well for lookup-style questions and fails quietly on questions whose answers live in the relationships between entities rather than in any single passage. Drawing on the recent literature on graph-augmented and agentic retrieval, and on patterns observed across enterprise solution-architecture engagements, this paper offers practitioners a workload-first decision framework: a taxonomy of five question shapes that flat retrieval cannot answer reliably, a five-level maturity ladder for graph-aware retrieval, a reference architecture, an evaluation approach that avoids common metric pitfalls, and a phased adoption roadmap.