TechPrufer
Get started
Exa
Exa logoExa · Feature trail

Academic Publication Search

9 posts · Aug 2023 – Jul 2026

What this trail covers

Exa has launched a dedicated search capability for academic publications, indexing approximately 350 million publications and 30 million authors. This system allows for semantic search using natural language, even with vague or incomplete queries, and has demonstrated superior recall and latency compared to existing academic search engines. The ingestion pipeline includes OCR and document parsing for PDFs, combined with metadata such as authors, citations, and publication history. The system also includes a company search benchmark, which fine-tunes retrieval models for company search and builds an optimized ingestion pipeline for entity matching. An ~800 query benchmark and open-sourced evaluation harness were created to measure progress. The benchmark includes retrieval and RAG tracks, with explicit handling for static vs. dynamic attributes.

All Exa trails
  1. SOTA Search Over Academic PublicationsThis post details the launch of Exa's dedicated search capability for academic publications. It introduces a new index of 350 million publications and 30 million authors, enabling semantic search via natural language. The ingestion pipeline uses OCR and document parsing for PDFs, integrating metadata like authors, citations, and publication history. Evaluation benchmarks show Exa achieving 86.4% recall and 0.578s mean latency, outperforming competitors. The system combines web index and publication index results for reranking.
  2. How Search Quality Shapes RL OutcomesThis 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.
  3. Introducing Exa's Company Search BenchmarksThis post introduces Exa's Company Search Benchmarks, detailing the creation of an ~800 query dataset and an open-sourced evaluation harness. It highlights the challenges of evaluating company search, particularly distinguishing retrieval from LLM memorization, and the need for fresh, structured data. The benchmark includes retrieval and RAG tracks, with specific examples of queries and constraints for attribute filtering, dynamic facts (employee count, funding), and composite queries. It also addresses the temporal nature of dynamic facts and the importance of index freshness. This contributes to the broader goal of building an evaluation ecosystem for entity types and search domains.
  4. Introducing Exa's People Search BenchmarksIntroduced a new benchmark for people search, including a role-based people search dataset (1,400 queries) and an evaluation harness. Developed a retrieval system for people search, indexing over 1 billion people using hybrid retrieval (fine-tuned Exa embeddings + Exa search). The ingestion pipeline is designed to handle over 50 million updates per week. Analyzed 10,000 historical people queries to identify dominant usage patterns (role-based search, skill/role-based discovery, specific individual lookup). Generated role-based queries using a structured taxonomy and sampled executives/founders across different company market cap tiers. Implemented evaluation methodology using recall@k, NDCG for targeted lookups, and an LLM judge for role-based discovery.
  5. Our AI Research: How We Evaluate Semantic Search TechnologyThis post details Exa's methodology for evaluating semantic search technology, specifically for academic publications. It introduces two primary evaluation approaches: pure result grading using LLM graders on various query sets (in-the-wild, MS Marco, Exa Olympiad) and RAG grading using LLM agents with search results for question answering. The post also outlines Exa's evaluation philosophy, advocating for downstream task integration and granular feedback, and critiques traditional closed evaluation methods like MS Marco for their limitations regarding false negatives, scale, and distribution mismatch. The core contribution is the technical depth on how Exa measures and iterates on its search engine's quality, particularly for academic content.
  7. How We're Building the Next Generation of Search with Semantic Search TechnologyThis post details Exa's foundational infrastructure for next-generation search, emphasizing the shift from keyword-based search to an embedding-based neural search algorithm. It highlights the three core components: crawling (distributed network, custom HTML parser, S3 storage), preprocessing (training specialized transformer models to generate document embeddings), and serving (building a custom vector database). The post introduces the 'Exacluster' (144 H200 GPUs) as a testament to the 'Bitter Lesson' philosophy, enabling extensive model training. It also outlines two product offerings: Exa Search (low latency) and Exa Websets (high-computation, long-latency queries), with a future vision of merging them into a 'GPT-5 search model'. The post implicitly contributes to the academic publication search by describing the underlying technology that enhances its semantic search capabilities and overall performance.
  8. Perfect Web Search for AI Agents with Semantic Search TechnologyThis 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.
  9. Announcing Exa: The AI Search Engine with Semantic Search TechnologyThis 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.
  10. Building Next-Gen Search with Semantic Search TechnologyThis 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.