JulianFlux: Electrodynamic vector retrieval and causal routing for AI

JulianFlux presents an alternative retrieval and routing framework for AI agents that replaces static Euclidean vector databases with a continuous electrodynamic field. Rather than using distance metrics like cosine similarity, embeddings are projected from high-dimensional semantic space into a 128D manifold via the Julian-Gauss Fast Transform, and each vector is given a dynamic Gaussian bandwidth based on nearest-neighbor density. Data items receive topological charges (truth +1, contradiction −1) via lightweight geometric-median clustering or an NLI cross-encoder; a modified Lorentz Force (F = Q(E + S)) then routes agents by combining a repulsive electric field that blocks hallucinations, a decaying induction term that phases out deprecated facts, and a Poynting flux computed with geometric algebra that generates semantic momentum to push agents through causal step sequences.

Practical Python implementations demonstrate the claims: euclidean_baseline.py shows standard RAG failure on poisoned datasets, julianflux_educational_mvp.py runs a fast Lorentz-Force-based engine, scale_ab_benchmark.py performs adversarial large-scale A/B tests (reporting a 0.0% hallucination rate), and julian_flux_engine.py upgrades truth verification with a DeBERTa-v3 NLI model and temporal decay. The project positions a Rust/CUDA enterprise core for production, arguing that zero-copy multivector physics at low latency scales the approach across billions of vectors.