- Published on
JulianFlux: Electrodynamic vector retrieval and causal routing for AI
- Authors
- Name
- DeepHorus
JulianFlux has open-sourced its core mathematics in Python, introducing an electrodynamic vector retrieval and causal routing engine designed to fix autonomous AI agents in enterprise environments. The project argues that standard vector databases such as Pinecone and Milvus function as static filing cabinets, relying on discrete Euclidean metrics like cosine similarity. Those systems blindly retrieve contradictory information, causing hallucinations and trapping agents in logic loops. JulianFlux proposes replacing distance measurement with continuous field physics.
The architecture compresses high-dimensional embeddings into a low-rank manifold. Its Julian-Gauss Fast Transform projects 768D semantic spaces down to a mathematically sound 128D manifold using a normal distribution. This preserves mapping accuracy while drastically reducing arithmetic cost. Dynamic bandwidths calculate Gaussian bandwidth for each vector based on nearest-neighbor distance, allowing the field to adapt to local semantic densities.
At the core, JulianFlux uses a modified Lorentz Force equation to route agents. An electric field mathematically repels hallucinations by assigning topological charges to data via NLI Logic Gates: −1.0 for contradictions and +1.0 for truth. The repulsive electric field physically blocks AI from accessing poisoned data. Faraday's Law of Induction decays old facts and corporate policies over time, generating a repulsive field against deprecated workflows.
The Poynting Flux extracts causal sequences to create Semantic Momentum. By computing the interior contraction product in Geometric Algebra, S = E ⌟ B, the system generates a momentum vector that pushes an agent sequentially from Step A to Step B to Step C. This transforms retrieval into kinetic reasoning. The company says continuous field architecture outperforms standard Euclidean RAG.
JulianFlux released four local Python files for validation. The first, euclidean_baseline.py, demonstrates standard cosine similarity retrieval on a poisoned dataset, showing RAG failing and collapsing causal workflow. The second, julianflux_educational_mvp.py, provides a lightweight engine with 128D projection, dynamic σ, and Lorentz Force, using fast Geometric Median clustering for truth charges. The third, scale_ab_benchmark.py, runs an adversarial scale test with thousands of background vectors, proving a 0.0% hallucination rate at scale. The fourth, julian_flux_engine.py, replaces fast clustering with a heavy DeBERTa-v3 NLI Cross-Encoder for enterprise-grade truth verification, adding Faraday's Law and Temporal Decay.
The open-source repository is available via git clone at github.com/julianjohnson-web/julianflux-research. Users install dependencies with pip install -r requirements.txt and run the scale benchmark with python scale_ab_benchmark.py. The Python bindings are intended for educational research and mathematical validation. For production, JulianFlux Enterprise Cloud executes the multivector physics equations in a zero-copy Rust/CUDA backend.
That enterprise backend eliminates the Python Global Interpreter Lock, allowing Lorentz Force equations to execute across billions of vectors with sub-millisecond p99 latencies. The company is seeking Enterprise Cloud access inquiries and design partnerships. The release signals a potential shift from static vector search to dynamic, physics-based routing for AI agents. It remains to be seen how widely enterprises adopt this approach.
