Academic Publication Search

7/23/2026

This post introduces a new, dedicated search system for academic publications. It details the scale of the index (350M publications, 30M authors), the semantic search approach that handles natural language queries and imperfect recall, and introduces new evaluation benchmarks (known-item and tip-of-the-tongue retrieval). It also outlines the technical components of the ingestion pipeline (OCR, document parsing) and the query-time process (combining web and publication index results). Performance metrics show Exa outperforming Google Scholar, Perplexity, and other systems in recall and MRR.

5/13/2026

This post details an experiment comparing Exa's search backend with Google's SERP for training RL agents. It highlights how using Exa as the retrieval backend during RL training leads to higher performance with less training compute. The experiment involved training Qwen3-4B-Instruct-2507 models with LoRA adapters, using a consistent system prompt and tool configuration (5 live web results, 2000 character snippets). The analysis focuses on the impact of the retrieval backend on agent performance, token efficiency, and the trade-offs between search engine performance during training versus inference.

1/22/2026

Introduced a new benchmark for company search, including an ~800 query dataset and an open-sourced evaluation harness. This benchmark focuses on retrieval and RAG tracks, with specific handling for static vs. dynamic company attributes (founding year, location, industry, employee counts, funding history). The dataset was designed to avoid well-known companies to test retrieval capabilities rather than LLM memorization. The post details query types for attribute filtering, funding, composite, and semantic queries, as well as RAG queries for extracting static and dynamic facts.

12/17/2025

This post introduces Exa's new people search capability, which indexes over 1 billion people using hybrid retrieval combining fine-tuned Exa embeddings with Exa search. The ingestion pipeline is designed to handle over 50 million updates per week. The post details the methodology for generating role-based queries and discovery queries, including sampling executives from various company tiers and using an LLM for structured taxonomy generation. It also outlines the evaluation methodology using recall@k, NDCG, and an LLM judge for role-based discovery. An open-source evaluation harness and dataset are released as a public benchmark for people search.

5/30/2025

This post details Exa's evaluation methodology for its semantic search technology, including LLM-based pure result grading and RAG grading. It also outlines a philosophy for search evaluation that prioritizes downstream task performance and granular feedback, critiquing traditional closed evaluation methods like MS Marco for their limitations regarding false negatives, scale, and distribution mismatch. The post highlights Exa's own evaluation results, showing state-of-the-art performance on various query sets, particularly on challenging Olympiad datasets.

1/7/2025

This post details the vision and technical requirements for a 'perfect search engine' that goes beyond keyword matching to understand precise meaning, enabling semantic search for AI agents. It outlines capabilities for searching by idea, people, multimodal content, and with perfect control and comprehensiveness. The post emphasizes the need for a search engine that can handle complex queries and provide exact results, contrasting it with current keyword-based systems and AI summaries that rely on them. It highlights the potential impact on scientific progress, community building, recruiting, and overall information access.

7/12/2024

This post articulates the vision and necessity of 'superknowledge' as a foundational element for future AI development, contrasting it with 'superintelligence'. It highlights the limitations of current knowledge systems (like Google) and positions Exa's academic publication search as a step towards achieving this superknowledge. The post emphasizes the technical challenge of building systems capable of handling complex retrieval requests and argues that Exa's usage-based revenue model and focus on novel ML architectures are key differentiators in pursuing this goal.

1/25/2024

This post announces the rebranding of Metaphor to Exa and introduces 'highlights' as a new feature. Highlights allow for instant extraction of webpage content from search results using customizable embedding models, achieved by chunking and embedding full webpages with a paragraph prediction model. This feature is designed to provide highly relevant and dense context for AI applications, improving LLM performance by filtering out noisy SEO results and retrieving specific content based on meaning rather than keywords. The post also reiterates Exa's core mission of organizing the world's knowledge through semantic search powered by transformers, contrasting it with traditional keyword-based search engines.

8/8/2023

This post discusses the evolution of Exa's search technology, highlighting its application for LLMs. It details how Exa's search engine, originally built for humans, is also suitable for LLMs due to its quality focus, high precision, LLM-native query understanding, content retrieval capabilities, low latency, and content moderation. The post emphasizes the synergy between LLMs and Exa's search API for powering AI applications.