Show HN: RIHU – Beyond RAG: Retrieval as Geometry, Not Vectors (github.com)

RIHU is an open-source prototype retrieval system built on a new paradigm called KAG (Knowledge as Geometry), which models knowledge as points, regions and relations in a continuous geometric “Universe” rather than as single-purpose embedding vectors. The repo demonstrates retrieval-as-navigation: queries become vantage points in a multi‑axis coordinate space (time, place, semantics, etc.), and relevance is determined by spatial relations — distance (aboutness), density (salience) and containment (context). RIHU supports both objective (world‑centric) and subjective (observer‑centric) searches, offers interactive visualizations, and is intended as a research artifact for explainable, bias‑transparent retrieval and hybrid RAG pipelines. Technically, KAG represents Hypothetical Facts (HFs), Classes and Instances as coordinates. RIHU’s construction pipeline is LLM‑assisted for suggesting axes and extracting HFs, then initializes coordinates via embeddings + SVD, optionally aligns anchors with a one‑shot affine transform, and refines positions with iterative relaxation. Key metrics are centroids (mean of instance coordinates), influence radius/enclosing ball, volume (r^D), and power = instances / volume — enabling notions like class subordination (A inside B) and contextual containment (HF inside class influence). The Python 3.9 prototype (sample using Sherlock Holmes text) outputs universe.json/axes/searches and demonstrates how geometric semantics yield interpretable, navigable retrieval that can integrate with existing vector methods.