In-Browser Agents for Search Assistance

TL;DR

The paper tackles the privacy-risk trade-off in AI-assisted web search by proposing a fully client-side browser extension that combines an adaptive probabilistic user policy with an in-browser Small Language Model (SLM) to generate context-aware suggestions. The system runs entirely on the user’s device, grounding the SLM with a locally learned policy and executing inference via WebGPU, thereby preserving data sovereignty. It introduces three modules—Behavioral Observation, Dynamic User Modeling, and Hybrid Inference—and demonstrates online personalization through REINFORCE-style updates, achieving significant gains in next-action prediction accuracy and search efficiency in a 3-week longitudinal study with 18 participants. The study reports high usability and trust scores, a substantial reduction in session length, and a noteworthy acceptance rate for suggestions, underscoring the practical viability and privacy benefits of the approach. The work contributes a novel, open-source, privacy-centric framework for in-browser AI-assisted search and shows that sophisticated, personalized AI assistance can be achieved without centralized data collection.

Abstract

A fundamental tension exists between the demand for sophisticated AI assistance in web search and the need for user data privacy. Current centralized models require users to transmit sensitive browsing data to external services, which limits user control. In this paper, we present a browser extension that provides a viable in-browser alternative. We introduce a hybrid architecture that functions entirely on the client side, combining two components: (1) an adaptive probabilistic model that learns a user's behavioral policy from direct feedback, and (2) a Small Language Model (SLM), running in the browser, which is grounded by the probabilistic model to generate context-aware suggestions. To evaluate this approach, we conducted a three-week longitudinal user study with 18 participants. Our results show that this privacy-preserving approach is highly effective at adapting to individual user behavior, leading to measurably improved search efficiency. This work demonstrates that sophisticated AI assistance is achievable without compromising user privacy or data control.