papers/hanzo-search/sections/01-introduction.tex

\section{Introduction}

\label{sec:intro}

Web search has undergone a fundamental transformation with the advent of large language models. Traditional search engines return ranked lists of documents, requiring users to click through multiple results, read pages, and synthesize information themselves. Generative search systems~\cite{metzler2021rethinking,nair2023bing} promise to deliver direct answers, but face critical challenges in factual accuracy, source attribution, and answer format.

\subsection{Challenges in Generative Search}

\begin{enumerate}[leftmargin=1.4em]
    \item \textbf{Hallucination}: LLMs generate fluent but factually incorrect content, with hallucination rates of 15--30\% in naive RAG implementations~\cite{huang2023hallucination}.
    \item \textbf{Attribution}: Users need to verify claims against sources, but current systems provide coarse document-level citations rather than sentence-level provenance.
    \item \textbf{Format rigidity}: Text-only answers fail to leverage the expressiveness of modern web interfaces for presenting structured information.
    \item \textbf{Retrieval quality}: The quality of generated answers is bounded by the quality of retrieved context. Poor retrieval cascades into poor generation.
\end{enumerate}

\subsection{Our Approach}

Hanzo Search addresses these challenges through a three-stage pipeline:

\begin{enumerate}[leftmargin=1.4em]
    \item \textbf{Hybrid Retrieval}: Fuse sparse and dense retrieval signals with learned weights, enhanced by query expansion and re-ranking.
    \item \textbf{Grounded Generation}: Generate answers with inline source citations, enforced by a constrained decoding scheme that requires every factual claim to be grounded in retrieved passages.
    \item \textbf{Generative UI}: Render answers as dynamic, interactive components tailored to query intent.
\end{enumerate}

\subsection{Contributions}

\begin{enumerate}[leftmargin=1.4em]
    \item A hybrid retrieval pipeline with learned fusion achieving state-of-the-art NDCG@10 on BEIR (\S\ref{sec:retrieval}).
    \item A constrained generation framework ensuring sentence-level source attribution (\S\ref{sec:generation}).
    \item A generative UI system with 12 component types driven by intent classification (\S\ref{sec:genui}).
    \item Comprehensive evaluation on standard and novel benchmarks (\S\ref{sec:evaluation}).
    \item Production analysis from 2.1M queries over 12 months (\S\ref{sec:production}).
\end{enumerate}