Abstract
Retrieval-augmented generation was supposed to ground language models in an organization's own knowledge. In practice, the dominant implementation destroys that knowledge on the way in. Standard RAG pipelines fragment structured content into fixed-size chunks, discarding the hierarchies, types, and relationships that organizations spent decades encoding — and then ask a vector index to recover, by statistical similarity, the meaning the chunker just threw away.
This paper describes Structured Context Retrieval (SCR), a retrieval methodology built on a different premise: structure is meaning, and retrieval should preserve it rather than reconstruct it. SCR retrieves components — semantically meaningful units defined by their schemas, with type information and relationships intact — instead of chunks, arbitrary fragments defined by token counts. We define the component–chunk distinction, state the architectural commitments SCR entails, locate SCR within the current Graph RAG discourse by distinguishing schema-grounded graphs from LLM-extracted ones — a distinction that reverses the standard cost argument against graphs — and present the converging evidence: schema-aware retrieval reaching 92% accuracy against a 68% vector-only baseline, hallucination reductions on the order of 90%, and independent 2026 results that replicate the pattern. SCR is the retrieval discipline implied by the broader principle of consulting authority before generating — look up before you make up — which we develop in a companion paper.
