DigUp is a free Mac app that searches files by their content using Google DeepMind's EmbeddingGemma 2 model, enabling users to find photos, PDFs, documents, audio and video across 100+ languages through natural language queries. All processing happens locally on the Mac with no data leaving the device.
BreadBowl-Embed proposes a new retrieval representation that sits between single-vector and token-level approaches, using 16 fixed slots per passage with dual vectors for routing and value reading. This addresses the inefficiency of two-stage RAG pipelines where documents are read twice—first for retrieval via bi-encoder, then for reranking via cross-encoder—reducing computational waste especially for agents that issue multiple queries.
A novel search reranking technique treats queries as bags of decisions by generating LLM-based yes/no questions specific to each query, then scoring document relevance by summing the probability of affirmative answers from Jev, a reranker model. Experiments on e-commerce datasets (Wayfair WANDS and Amazon ESCI) show this multi-question approach outperforms BM25 and single-question reranking.
Researchers measured how AI-search platforms select and cite sources, finding that citations concentrate in platform-specific domains with low publication barriers. They demonstrated that ordinary posts on preferred platforms can be cited within days, and that commercial services can exploit this vulnerability to inject content into AI-search results.
A reranking technique that generates query-specific relevance criteria as yes/no questions instead of traditional bag-of-terms approaches. An LLM produces a rubric of decision questions for each query, then Jev measures document relevance by summing the probability of affirmative answers across all questions. Experiments on e-commerce datasets show this bag-of-decisions method outperforms BM25 and single-question relevance scoring.
Sup is a console-based email client designed for managing large volumes of email through a threaded interface with tagging support. It offers fast full-text search, multi-account handling, custom Ruby extensions, GPG encryption, and organizational features like labels and automatic contact management.
A writer experimented with using LLMs to generate personalized book recommendations by having detailed conversations about literary preferences, rather than relying on search history alone. The approach showed promising results and led to a prototype app that stores user preference profiles locally and uses conversational analysis to build evolving recommendation models, potentially offering privacy benefits over traditional recommendation systems.
Eniac is a web search API that converts millions of web pages into cited datasets. Users can decompose queries, filter results in real-time with cost tracking, manage API keys for programmatic access, and maintain a billing wallet with top-up functionality.
SearXNG is a decentralized web search engine that aggregates results from automated crawls and the YaCy peer-to-peer network, with the instance maintainer disclaiming responsibility for content distributed by other peers.
Microsoft is overhauling Windows Search with a faster, more modern redesign built on WinUI 3, featuring improved performance, better typo handling, and new capabilities like action commands and inline results. The new search interface is currently in testing with Windows Insiders and will roll out to Windows 11 in coming months.
Microsoft is rebuilding Windows 11's search functionality to include typed commands and integrated Copilot features directly in the search interface.
Anthropic previews ATLAS, a new benchmark for evaluating AI agents on search-intensive tasks with 547 real-world research queries paired with verified answers. The benchmark reveals that comprehensive web search remains costly and incomplete, with even top agents missing about one-third of relevant results, and no agent under $1 per task achieved strong performance metrics.
Google's search engine centralized internet discovery but created perverse incentives—paid results, clickbait, and algorithmic gaming degraded content quality. AI companies now propose similar solutions to this self-inflicted problem, but they merely summarize the same polluted internet.