Engineering the Topology of Knowledge
A four-article series on why RAG fails and how to fix it
Naïve RAG — chunk, embed, cosine retrieve, concatenate, generate — is Abbott's Flatland inhabitant: it queries a cosine projection and mistakes the shadow for the sphere. The knowledge object above the projection plane is typed, relational, temporal, hierarchical. The projection discards all of it.
This series names the substrate problem, diagnoses four ways the field has mistaken shadow for object, and points the engineering effort downward — toward ingestion, not orchestration.
The collision geometry of dense embeddings. Why cosine similarity cannot separate documents with opposite clinical outcomes. The Flatland metaphor. The dentate gyrus and sparse coding. Hyperbolic embedding. Four fallacies named.
Multi-modal extraction. Ontology-aware typed relation extraction. Provenance topology and source monitoring. Quality gates and corpus auditing. A reference ingestion pipeline from document to structured knowledge graph.
When the knowledge lives in multiple systems under different regulatory regimes. Privacy-preserving graph construction. Cross-domain identity resolution. The governance layer for multi-party knowledge engineering.
A production-ready ingestion-first RAG system. Benchmarked against the three diagnostic tests from Article I. Open-sourced. Battle-tested on a 10,000-document pharmaceutical corpus under regulatory constraints.