Overview
Exa (formerly Metaphor) represents a paradigm shift in web indexing, moving from traditional keyword-based crawling to a fully neural architecture. Unlike Google, which focuses on human-readable snippets, Exa is built specifically for Large Language Models (LLMs) and AI agents. It utilizes a transformer-based encoder to map the entire web into a high-dimensional vector space, allowing users to search by meaning and style rather than exact string matches. This enables 'find similar' capabilities that can locate documents with identical semantic structures across disparate domains. As of 2026, Exa has positioned itself as the essential retrieval layer for RAG (Retrieval-Augmented Generation) pipelines, offering clean Markdown outputs that bypass the noise of traditional HTML. Its architecture handles the heavy lifting of content scraping, boilerplate removal, and embedding-based filtering, providing a 'Search-as-a-Service' model that scales from hobbyist developers to enterprise-grade AI agents requiring high-concurrency access to the live web.