Exa.ai

Exa.ai builds web-search infrastructure for AI applications, combining its own crawling, retrieval models, search index, and page-content extraction. Formerly called Metaphor, the company offers a search API and Websets, a product for assembling structured lists from the web. Its consequential shift came after ChatGPT's release: demand for programmatic search turned a search engine built for people into an information source for AI products and agents.[2][7][3]

That AI focus followed earlier work on search by CEO and cofounder Will Bryk and his cofounder Jeff, including a crowdsourced-link search engine they built together as Harvard roommates.[4]

Exa began in 2021. Bryk recalls entering Y Combinator that summer with a pitch of “Google 2.0”, arguing that GPT-3's ability to understand nuanced language could be applied to search. The team raised seed funding, bought a GPU cluster, and spent more than a year experimenting with architectures and training datasets.[7][1][3]

The first search engine launched in November 2022, two weeks before ChatGPT, according to Bryk. The team initially worried that direct answers from language models would make its work less important. Requests for API access supplied a different direction: AI applications needed external information to address hallucinations, stale knowledge, and limits on what model weights could retain. Bryk dates the focus on an AI search API to early 2023.[7][3][8]

On January 25, 2024, Metaphor became Exa and launched highlights, which extracted relevant passages from search results. That generation of highlights chunked and embedded full webpages with a paragraph-prediction model. Users could specify passage length, the number of passages per page, and a secondary query selecting the content they wanted. The API also supported long queries and page-content retrieval for downstream processing.[2]

Exa's initial training objective used the text surrounding hyperlinks to predict the linked document. In Bryk's example, the model saw a description such as a cool aerospace startup while the following link was hidden; repeating this task taught it to retrieve documents from descriptions. He clarified that the model was retrieving the document rather than generating a memorized URL string: “a more accurate name could be document prediction”. Exa subsequently supplemented this base objective with synthetic data and supervised fine-tuning.[1][3]

Bryk calls this objective “a neural PageRank”: shared links provide a signal of importance, while the surrounding language teaches the model different ways people describe the same document. His example is a Paul Graham fundraising essay referred to in many different ways. The analogy explains Exa's approach to relevance and quality; its behavior also depends on the training material the team selects.[1]

The search stack discovers URLs, crawls pages, processes documents into an index, and serves matching results. Returning page text became an important part of the product because AI customers needed cleaned content as well as links. Bryk recalls retaining the first 1,024 tokens of pages for debugging in the earliest version; that storage made it easier to respond when customers began asking for content.[1]

Exa also used LLMs to label search relevance. Bryk describes an internal exercise in which he and engineers struggled to judge niche queries, including biological topics, and fell back on keyword overlap. He said in January 2025 that the company was using LLM labelers to interpret those matches.[1]

Websets extended Exa toward structured lists of people, companies, and other entities. Rather than treating every search as a brief interactive request, it allowed more compute to be spent gathering results. In the January 2025 interview, Bryk described previews that users could inspect before launching larger jobs, with searches potentially taking minutes or longer.[1][7]

Bryk distinguished two problems: defining what qualifies as a match and finding enough matching results. A request for hardware startups in San Francisco, for example, could include robotics or wearables companies without those pages using the word hardware. Previews helped users refine their criteria; larger jobs pursued broader coverage. Finding every matching entity was the product's ambition, rather than an established guarantee of completeness.[1]

Exa built an in-house vector database to search document embeddings while supporting filters. Its December 2024 implementation account describes Matryoshka training, which makes shorter prefixes of an embedding useful representations in their own right. The system shortened 4,096-dimensional embeddings to 256 dimensions and compressed document components to one-bit values, while keeping query components as floating-point numbers.[5]

For each query subvector, the database precomputed dot products against possible binary document subvectors and stored lookup tables in CPU registers. Exa used C++ for this performance-critical section. It grouped documents into 100,000 similarity clusters, searched selected nearby clusters, and fetched extra candidates for reranking with uncompressed data to recover recall lost through compression. Inverted indexes restricted searches to documents matching date, domain, keyword, or other filters.[5]

Exa reported searching billions of vectors in under 100 milliseconds with high recall and costs ten times below quotes from cloud vector-database services. These figures concern the database implementation and quoted alternatives; the same article lists more than 500 queries per second as an engineering requirement.[5]

Owned GPUs were an early investment. Technical staff member Carlos Marques wrote in May 2025 that Exa had spent half its 2021 seed money on ten nodes containing 80 A100 GPUs. A subsequent $5 million purchase added an 18-node fleet with 144 H200 GPUs, bringing the two fleets to 224 GPUs. Bryk had already described the H200 purchase in the January 2025 interview, alongside a mix of owned infrastructure and AWS inference.[6][1]

Marques credited Pulumi, Ansible/Kubespray, and NVIDIA's operators with provisioning and configuring the fleet. Alluxio combined local NVMe storage into a distributed cache while S3 remained the authoritative store. Flyte scheduled reproducible ML workflows, including distributed training, checkpointing, dataset lineage, and cloud overflow. The workloads included embedding billions of pages and training a reranker.[6]

By May 2026, Exa described specialized embedding models fine-tuned for public code and technical-documentation search, alongside people and company indexes for go-to-market agents. Its research program also expanded into coding-agent search evaluations through WebCode, search-pipeline orchestration through Canon, agentic search endpoints, academic retrieval, and adaptive highlights.[8][9]

• July 16, 2024: Exa announced $22 million in combined seed and Series A funding led by Lightspeed Venture Partners, with participation from NVentures and Y Combinator. Latent Space separately reported the Series A as $17 million. Exa reported thousands of company and developer integrations and revenue tripling over the preceding few months, without an absolute revenue figure.[4][1]
• September 3, 2025: Exa announced an $85 million Series B led by Benchmark at a $700 million valuation, with Lightspeed, Y Combinator, and NVentures participating. Benchmark's Peter Fenton was to join the board. Planned investments included expanded indexing, a five-times-larger GPU cluster, and team growth.[7]
• May 20, 2026: Exa announced a $250 million Series C led by a16z at a $2.2 billion valuation. It planned next-generation models and infrastructure capable of handling hundreds of thousands of searches per second, and announced that LaunchDarkly president Marcus Holm was joining as chief revenue officer.[8]

In the May 2026 announcement, Exa reported more than 400,000 developers and over 5,000 companies using its service, naming Cursor, Cognition, HubSpot, OpenRouter, and Monday.com. It also reported crawlers tracking over 500 billion URLs. The usage figures did not specify active or paying accounts, and tracked URLs were distinct from indexed documents. The announced throughput expansion remained a capacity target.[8]