TechPrufer
Get started
Exa
Exa logoExa · Feature trail

Exa Instant

1 post · Dec 2024 – Feb 2026

What this trail covers

Exa 2.0 significantly enhances search capabilities with Exa Fast (sub-425ms P50 latency), Exa Auto (optimized balance of latency and quality), and Exa Deep (agentic retrieval for highest quality results). The underlying infrastructure includes a larger index (tens of billions of webpages refreshed every minute), a fine-tuned embedding model trained on a 144x H200 cluster using new embedding techniques, and a significantly updated in-house vector database with new clustering algorithms, lexical search optimizations, and advanced memory/speed optimizations including Matryoshka embeddings, binary quantization, optimized dot-product calculations with lookup tables, hardware register utilization, and clustering for efficient search. The system also supports efficient metadata filtering via inverted indexes and reranking for recall recovery.

All Exa trails
  1. Introducing Exa Instant: the fastest search engine in the worldThis post introduces Exa Instant, a new search API endpoint designed for sub-200ms latency, significantly faster than previous offerings and competitors. It highlights the optimization of the entire search and retrieval stack to achieve this speed, contrasting it with slower APIs that wrap Google SERP. The post emphasizes the value of low-latency search for AI agents and real-time applications, detailing performance benchmarks and quality evaluations. It also mentions the availability of higher-tier options like Exa Fast, Exa Auto, and Exa Deep for different use cases.
  2. Introducing Exa 2.1Exa 2.1 announces significant improvements to its search API endpoints (Exa Fast, Exa Auto, Exa Deep) by scaling pre-training and test-time compute by an order of magnitude. Exa Fast achieves sub-500ms latency with high accuracy due to an independent document index and retrieval algorithm. Exa Deep offers the highest accuracy via agentic multi-step searches. The MCP quality is enhanced with a new deep search tool. The post details evaluation methodology using GPT-5 mini and gpt-4o-mini for grading.
  3. Introducing Exa 2.0This post details the technical advancements behind Exa 2.0, focusing on the infrastructure and algorithms that enable Exa Fast (latency improvements), Exa Auto (quality/latency balance), and Exa Deep (agentic retrieval). Key contributions include: a larger and more frequently refreshed web index, a new embedding model trained on a large GPU cluster with novel techniques, and substantial optimizations to the in-house vector database (clustering, compression, Rust assembly optimizations). It also highlights the open-sourcing of latency benchmarking code and the use of specific LLM evaluation harnesses for search quality.
  4. The World's Fastest Search APIIntroduces Exa Fast, a streamlined search API with p50 latency below 425ms. This is achieved by building the search engine from scratch, including crawling, model training, and an in-house vector database, allowing for end-to-end latency optimization. Benchmarks show Exa Fast is over 30% faster than competitors like Brave and Google Serp, which are often wrappers. The post emphasizes the importance of low latency for AI systems, agentic workflows, and latency-sensitive AI products.
  5. How We're Building the Next Generation of Search with Semantic Search TechnologyThis post details the foundational infrastructure for Exa's next-generation search, including a distributed crawling system, a custom HTML parser, S3 storage, and a proprietary vector database optimized for serving embeddings. It emphasizes the use of neural models for document preprocessing into embeddings, contrasting this with traditional keyword-based search. The post also introduces two product offerings: Exa Search for low-latency general search and Exa Websets for computationally intensive, high-quality result retrieval.
  7. Web Search API Evals: Exa's Neural Network Search Engine vs. The CompetitionThis post introduces new benchmarks (SimpleQA and MSMARCO) demonstrating Exa's state-of-the-art performance in web search for LLMs. It details the evaluation methodology, comparing Exa's search API with competitors like Bing and Perplexity, and highlights Exa's custom neural network search engine designed for LLM integration. The post quantifies improvements in factuality and retrieval accuracy, positioning Exa as a leading provider for RAG and classic search applications.
  8. How We Built a Web-Scale Vector Database for Our Neural Network Search EngineThis post details the development of Exa's web-scale vector database, focusing on five key optimizations: Matryoshka embeddings to reduce dimensionality by 20x, binary quantization to further reduce memory by 16x, a novel hybrid dot-product similarity metric using precomputed lookup tables for speed, maximizing hardware utilization by loading lookup tables into CPU registers, and clustering to improve throughput by ~1000x. It also describes the implementation of inverted indexes for efficient metadata filtering and a reranking step to recover recall.