\documentclass[10pt]{beamer}
\usetheme{Berkeley}
\usecolortheme{seahorse}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{amsmath,amssymb}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{hyperref}
\title[Neural Retrieval]{Neural Retrieval at Web Scale \\ Ideas, Trade-offs, and Open Problems}
\author{First Last}
\institute[Example U]{Department of Computer Science \\ University of Example}
\date{\today}
\begin{document}
\frame{\titlepage}
\section{Introduction}
\begin{frame}{Why Retrieval?}
\begin{itemize}
\item Retrieval powers search, RAG pipelines, and recommendation.
\item Classical lexical methods (BM25) remain hard baselines.
\item Dense methods enable semantic matching.
\end{itemize}
\end{frame}
\section{Background}
\begin{frame}{From Sparse to Dense}
\begin{itemize}
\item Sparse: BM25, TF-IDF --- exact term match.
\item Dense: bi-encoders producing vector embeddings.
\item Hybrid: late fusion of scores or first-stage lexical retrieval + reranker.
\end{itemize}
\end{frame}
\section{Method}
\begin{frame}{Our Approach}
A two-stage pipeline:
\begin{enumerate}
\item Lexical retrieval of top-$k$ candidates.
\item Cross-encoder reranking with a distilled model.
\end{enumerate}
Knobs: $k$ (recall vs. cost), quantization (latency vs. accuracy), pruning.
\end{frame}
\section{Results}
\begin{frame}{Benchmarks}
\begin{table}
\centering
\begin{tabular}{lcc}
\toprule
System & nDCG@10 & Latency (ms) \\
\midrule
BM25 & 0.42 & 8 \\
Dense bi-encoder & 0.48 & 22 \\
Ours (hybrid) & 0.56 & 35 \\
\bottomrule
\end{tabular}
\end{table}
\end{frame}
\section{Conclusion}
\begin{frame}{Takeaways}
\begin{itemize}
\item Hybrid still wins on quality per dollar.
\item Query-time efficiency matters as much as offline accuracy.
\item The next frontier: end-to-end distillation.
\end{itemize}
\end{frame}
\begin{frame}{Thank You}
\centering
\Huge Questions?\\[0.8em]
\normalsize [email protected]
\end{frame}
\end{document}

PDF Preview
Create an account to compile and preview