papers/embedding-7680/embedding-7680.tex

\documentclass[11pt,letterpaper]{article}
\usepackage{shared/zoocover}

% Packages
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{amsmath,amssymb,amsthm}
\usepackage{algorithm}
\usepackage{algorithmic}
\usepackage{graphicx}
\usepackage{hyperref}
\usepackage{xcolor}
\usepackage{booktabs}
\usepackage{tikz}
\usetikzlibrary{shapes,arrows,positioning,calc}
\usepackage{listings}
\usepackage{caption}
\usepackage{subcaption}
% geometry loaded by zoocover

% Theorem environments
\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{definition}{Definition}[section]
\newtheorem{example}{Example}[section]

% Custom commands
\newcommand{\Zoo}{\textsc{Zoo}}
\newcommand{\Zen}{\textsc{Zen}}
\newcommand{\DSO}{\textsc{dso}}
\newcommand{\LLM}{\textsc{llm}}
\newcommand{\MTEB}{\textsc{mteb}}
\newcommand{\RAG}{\textsc{rag}}

% Hyperref setup
\hypersetup{
    colorlinks=true,
    linkcolor=blue,
    citecolor=blue,
    urlcolor=blue
}

% Title and authors
\title{\textbf{High-Dimensional Embedding Spaces for Semantic Search: The Case for 7680 Dimensions}\\
\large Zen Reranker and Native Embeddings for Decentralized AI}

\author{
Antje Worring, Zoo Labs Foundation Research\thanks{Corresponding author: [email protected]}\\
\textit{Hanzo Industries Inc (Techstars '17)}\\
\textit{Lux Industries Inc}\\
\textit{Zoo Labs Foundation}\\
\texttt{[email protected]}
}

\date{January 2025}

\begin{document}

\zoocoverpage

\begin{abstract}
We present a systematic analysis of embedding dimensionality for semantic search in decentralized AI systems, culminating in the \textbf{Zen Reranker}---a specialized embedding model with native 7680-dimensional output. Current embedding models produce heterogeneous dimensionalities (768 to 8192), requiring costly alignment layers when combining embeddings from different models. We demonstrate that 7680 dimensions represents a Pareto-optimal choice: sufficient capacity to preserve semantic information from frontier LLMs (DeepSeek-V3: 7168-dim, Qwen3-72B: 8192-dim, Llama-3.3-70B: 8192-dim) while enabling efficient BitDelta compression (31.87$\times$). The Zen Reranker achieves 68.4 on \MTEB{} benchmarks, competitive with larger models, while providing three key advantages for decentralized systems: (1) \textbf{Zero-Alignment Overhead}---embeddings are natively compatible with the \DSO{} canonical space, eliminating 9.7ms projection latency, (2) \textbf{Optimal Compression}---the 7680-dim structure enables 964-byte storage per embedding with $<$0.5\% RMSE via BitDelta quantization, and (3) \textbf{Cross-Model Retrieval}---94.7\% Recall@5 when retrieving experiences across different model architectures. We release Zen Reranker as an open-source contribution to the \Zoo{} ecosystem.

\textbf{Keywords}: embedding models, semantic search, high-dimensional spaces, quantization, decentralized AI, reranking
\end{abstract}

\input{sections/01-introduction}
\input{sections/02-background}
\input{sections/03-dimensionality-analysis}
\input{sections/04-7680-dimension-optimality}
\input{sections/05-zen-reranker-architecture}
\input{sections/06-bitdelta-compression}
\input{sections/07-evaluation}
\input{sections/08-integration-with-zoo-network}
\input{sections/09-related-work}
\input{sections/10-conclusion}
\input{sections/99-acknowledgments}

\end{document}