Exa

Exa builds web search infrastructure for AI applications, combining its own crawler, retrieval models and serving systems with webpage-content extraction. Formerly Metaphor, it renamed itself Exa on January 25, 2024. Its products developed from a search engine into an API, Websets for extensive entity lists, and a research agent that returns structured information. Will Bryk is its CEO and cofounder.[3][1][5][10]

Exa's neural search work followed an earlier project with a different approach. In its July 2024 funding announcement, Bryk recalled that he and cofounder Jeff built a search engine while roommates at Harvard: “At the time, we thought crowdsourcing links would enable better search than Google.” That collaborative project preceded the company's later model-based retrieval work.[4]

Bryk's September 2025 account dates Exa's start to 2021 and its first search-engine launch to November 2022, two weeks before ChatGPT appeared. He recalled that developers soon requested API access for AI applications. His May 2026 account places the AI-focused search API announcement in early 2023.[5][6]

The team initially wondered whether direct answers from ChatGPT would make its search work less important. In its August 2023 explanation, it instead identified hallucinations, training cutoffs and limited memorized knowledge as reasons LLMs needed external retrieval. The resulting API direction emphasized long, nuanced queries, precise results and access to page contents.[7]

Exa's initial training objective used the text surrounding links on the web to predict the associated document. In the January 2025 interview, Bryk illustrated this with a sentence recommending a “really cool aerospace startup”: hide the following link, train the model to identify its destination, then use a similar description as a search query. This was document prediction, rather than generation of URL strings. By that interview, he said training had expanded to include synthetic data and supervised fine tuning.[1][7]

Bryk called the objective “neural PageRank” because it could learn both a frequently shared document's prominence and the different language people used to describe it. His example was a Paul Graham fundraising essay referenced in many ways. This was his explanation of the model's quality signal; selecting high-quality training sources was another way Exa could influence its results.[1]

Exa built crawling, AI processing and serving as separate parts of its search engine. The crawler finds and fetches pages; processing turns documents into searchable representations; the serving system retrieves matches at query time. Bryk explained that expensive processing must happen ahead of searches, with query-time models reranking a smaller candidate set rather than examining billions of documents afresh.[1]

The team's December 2024 vector-database account explains how it reduced retrieval cost. Matryoshka training made shorter embedding prefixes useful, allowing 4,096-dimensional vectors to be shortened to 256 dimensions. Binary quantization then replaced each document dimension's 16-bit floating-point value with a one-bit sign. Queries retained floating-point values, enabling dot-product scoring against compressed documents.[8]

For each four-element query segment, Exa precomputed dot products with all 16 possible binary document segments and loaded lookup tables into CPU registers. It grouped documents into 100,000 similarity clusters so searches could examine relevant clusters and nearby ones. The system retrieved extra candidates and reranked them with uncompressed data to recover recall lost through compression. Inverted indexes restricted searches by metadata and keyword filters. The team reported searching billions of vectors in under 100 milliseconds; that figure concerned vector retrieval, rather than the full API request.[8]

Page-content retrieval became useful to customers building AI applications that needed cleaned text alongside URLs. Bryk recalled that the first version had retained the first 1,024 tokens of page content for debugging, which made serving content easier when users began requesting it. Exa subsequently offered parsed content, including Markdown and chunked text.[1]

The January 2024 rebrand also introduced highlights: query-relevant excerpts extracted from search-result pages. Exa chunked and embedded full webpages with a paragraph-prediction model, letting users customize excerpt length, excerpts per page and a secondary query for the desired content. The company tied its API business to retrieval quality, stating that it had no advertising-based need to monetize attention.[3]

Websets extended search into structured lists of people, companies and other entities. In the January 2025 interview, Bryk described spending more compute on complex queries and showing previews before users launched larger jobs. Finding hardware startups in San Francisco was an illustrative task, not a measured completeness result. He distinguished clarifying what counts as a match from finding more of the matching entities. By September 2025, Exa identified the API and Websets as its two products and offered zero data retention across its product endpoints.[1][5]

Exa also progressed from search to an answers endpoint and then a deep research agent. In its June 30, 2025 collaborator case study, LangChain described a LangGraph implementation with a planner generating parallel research tasks and an observer retaining context across planning, reasoning, outputs and citations. Individual tasks received other tasks' cleaned outputs rather than their intermediate reasoning. The case study explicitly credited lessons from Anthropic's Deep Research architecture.[10]

The research agent tried reasoning from search snippets before requesting full page contents and used JSON schemas and function calling to produce structured API output. LangChain reported hundreds of research queries daily, with results taking 15 seconds to three minutes depending on complexity. Exa engineer Mark Pekala credited LangSmith's token observability: “The observability – understanding the token usage – that LangSmith provided was really important. It was also super easy to set up.” The collaborator linked that visibility to pricing decisions.[10]

Exa developed specialized code-search embeddings and a dedicated ingestion pipeline for fresh documentation, changelogs and issues with less navigation and page boilerplate. On March 23, 2026, Joshua Ahn, Gabriel Cammany and Nitya Sridhar announced the open-sourcing of WebCode, evaluations separating content extraction, in-document highlights, web retrieval and end-to-end coding tasks.[6][11]

WebCode distinguished answer correctness from groundedness—whether retrieved evidence actually contained the answer. This addressed cases where a model answered correctly from memory despite poor retrieval. Its extraction reference combined Browserbase-rendered screenshots and the DOM with a multimodal model. The released datasets contained 250 extraction URLs, 317 retrieval question-answer pairs and 33 coding tasks across nine languages. Modal sandboxes verified that reference solutions passed tests while empty stubs failed.[11]

On August 28, 2026, Exa released Dynamic Highlights as a research preview. Instead of selecting excerpts independently from each page, it concatenated retrieved documents for one model forward pass and allocated tokens across the whole set. This allowed more space for useful material and less for irrelevant or duplicate information. Exa reported an average 95% token reduction against full page content. Separately, at a 12,000-character budget, its single-turn tests reported a 40% token-efficiency gain over existing highlights and a 3.8% quality increase. Specified Exa Agent evaluations reported 30% lower agent token use and a 2.1% average quality gain.[12]

Exa's April 28, 2026 Google partnership announcement described Grounding with Exa Web Search in Private Preview and the Exa Agent's launch on Gemini Enterprise's Agent Marketplace. The grounding integration supplied fresh search results and relevant page excerpts to Gemini; the marketplace agent supported web research, similar-page discovery and URL-content retrieval.[13]

Owned compute was an early consequential investment. In his May 15, 2025 infrastructure account, technical staff member Carlos Marques said Exa spent half its 2021 seed money on an 80-A100 cluster. After inference needs and parallel research workloads increased by late 2024, it added a $5 million, 144-H200 fleet. The two clusters totaled 224 GPUs in that account and supported concurrent webpage embedding, reranker training and checkpoint workflows.[9]

Marques described Pulumi provisioning, Ansible/Kubespray setup and NVIDIA GPU and network operators beneath the research workloads. Alluxio combined local NVMe storage into a distributed cache while S3 remained the authoritative store. Exa selected Flyte for scheduling, reproducible runs, checkpointing, distributed training and cloud execution support. Bryk also described using both owned infrastructure and AWS inference.[9][1]

On July 16, 2024, Exa announced $22 million in combined seed and Series A funding, led by Lightspeed Venture Partners with NVentures and Y Combinator participating. The January 2025 Latent Space introduction separately reported the Series A itself as $17 million. In July 2024, Exa reported integrations by thousands of companies and developers, including Databricks research teams assembling training datasets, and said revenue had tripled over the preceding few months.[4][1]

On 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. It announced that Benchmark's Peter Fenton would join its board and planned to expand indexing, its team and GPU capacity. In a January 9, 2026 article, Jan-Erik Asplund reported Sacra's estimate that Exa reached $10 million in annualized revenue in September 2025, approximately 11 times its year-earlier level.[5][14]

On May 20, 2026, Exa announced a $250 million Series C led by a16z at a $2.2 billion valuation. It reported more than 400,000 developers, more than 5,000 companies using Exa and crawlers tracking more than 500 billion URLs, naming Cursor, Cognition, HubSpot, OpenRouter and Monday.com as search users. These represented developer reach, company usage and crawler coverage—not paying-customer or indexed-page counts. The funding was intended for new models and infrastructure capable of hundreds of thousands of searches per second, a future target. Exa also announced that LaunchDarkly president Marcus Holm was joining as CRO to lead its global go-to-market organization.[6]