I built RIHU (Retrieval In the Hypothetical Universe) while wondering: how could we represent the world we live in as data?
RIHU is a small open-source prototype that explores what retrieval might look like beyond RAG — when knowledge becomes geometry, not just vectors.
Instead of embedding similarity search, RIHU constructs a geometric “universe” of Hypothetical Facts, Classes, and Instances.
Each concept exists along interpretable spatial axes (like time, location, or semantics), and retrieval becomes wayfinding in that space.
Key ideas: - Geometry over similarity: proximity = relevance, density = salience, containment = context - Two retrieval modes: - Objective (global): find centroids and influence regions - Subjective (observer-centric): retrieve from a vantage point - Metrics: centroid, radius, volume, power (instances per volume)
RIHU comes with a small Python implementation and a demo universe built from The Adventures of Sherlock Holmes. You can explore it through sample code or visualize it in 3D.
This project is a small step toward more interpretable, observer-aware, and geometric forms of retrieval. Questions, criticism, and forks are all welcome!