A Demo of Interactive Thematic Data Collection on the Live Web

Abstract

Many emerging AI systems can search and reason over web content, but most are designed for task-bounded objectives such as answering a question or producing a report. In contrast, many real-world workflows require thematic web-data collection: identifying and gathering large numbers of topically relevant documents distributed across the live web. We present a demonstration of an interactive web-data collection system for thematic collection building, implemented on top of a multi-agent focused crawling architecture with human-in-the-loop relevance and steering feedback. The system decomposes data collection into cooperating agents for query generation, page relevance assessment, and frontier exploration. These agents operate within a dynamic re-seeding loop that periodically re-enters the web through newly generated search queries, enabling exploration beyond the currently reachable crawl frontier and into disjoint regions of the web graph. A browser-based interface exposes the discovery process to users, who can inspect intermediate results, correct relevance judgments, and provide steering feedback that directly influences system behavior. The system records query histories, crawl statistics, and discovered pages as exportable artifacts. This demonstration highlights how interactive, agent-driven focused crawling can support long-horizon thematic data collection workflows in which users aim to build reusable collections of web resources rather than obtain a single synthesized answer.