ExaTranslation site

Exa is a search infrastructure provided by Exa Labs Inc. that primarily serves developers and AI applications. It combines web search, relevant paragraph extraction, and research interfaces to provide applications with results, raw links, and textual evidence that can be further processed. Common uses include helping coding assistants look up documentation, supplementing Q&A systems with new information, and establishing research workflows for businesses or industries.Exa Product Description

What's new in Exa 3.0?

Exa released version 3.0 on October 19, 2026, focusing on upgrades to the retrieval model, content filtering, and index coverage, while simultaneously launching a more affordable “Lite” search mode. This is not a “3.0 client” that needs to be downloaded to a computer; existing API integrations will continue to use the service via the corresponding interfaces with the search mode.

This update mainly includes:

• Enhanced Search Capabilities: Larger embedding models are used, and the candidate result generation and ranking processes have been optimized.
• Dynamic Highlights is now officially available: it distributes excerpt lengths across multiple results to reduce duplicate content. According to the official report, it reduces the number of tokens by an average of approximately 95% compared to the full page text. This data is based on the evaluation conditions and does not represent a fixed savings ratio for all requests.
• Expand Data Coverage: Add or expand information sources for products, local businesses, news, weather, sports, and travel.
• Differences between Lite and Auto: Lite is designed for applications that are sensitive to cost and interaction latency; Auto uses more computational resources during the retrieval and ranking phases.

These updates should be verified by incorporating them into the app’s own query set; one should not rely solely on vendor rankings to infer search performance in a specific Chinese vertical, a particular country, or on a specific website.Exa 3.0 Release Announcement

Official Information:Official Website: https://exa.ai/; Version 3.0 Announcement: https://exa.ai/blog/exa-3; Pricing Information: https://exa.ai/pricing.

Core Features and Practical Value

Search results can be directly incorporated into the application process

The Search API returns web links and related content. Developers can request highlighted excerpts, full text, or other output formats, eliminating the need to “first retrieve search results and then write scraping logic page by page.” It also supports filtering by domain and narrowing down results by content type—for example, to find only official development documentation, news, or company information.Search API

Using Excerpts to Control Contextual Costs

Standard Highlights provide excerpts from the original text that are relevant to the query, while Dynamic Highlights further take into account the value and duplication among multiple results. This feature is useful for reducing irrelevant web content in multi-round agent workflows. Summaries or excerpts omit the page’s context; when eligibility, pricing, or exceptions are involved, you should still refer to the full source for more information.Highlights User Guide

Extract content directly when a link is provided

If you already know the target URL, there’s no need to perform another broad search. The Contents API can read page content and uses the cache age control to determine when to attempt to retrieve updated content. Exa’s crawler adheres to website crawling rules and does not guarantee the ability to bypass logins, paywalls, or CAPTCHAs.Product Selection and Crawling Boundaries

Expanding from a Single Search to a Research Workflow

Exa also provides an Agent API for handling asynchronous, multi-step studies, as well as Monitors for periodically searching for and receiving new results. When selecting these capabilities, costs should be estimated separately for each task; the per-query price for a standard Search cannot be directly applied to all study operations.Exa Developer Documentation

Who is this product suitable for?

• AI Application and Agent Developers: Provide the model with source links and relevant web content to generate traceable answers.
• Programming Tools Team: Retrieve the latest SDK documentation, project descriptions, and change logs to reduce the risk of relying solely on existing knowledge of the model.
• Content and Research Team: First, identify a set of candidate sources; then, conduct a manual review or secondary curation.
• Enterprise Data Product Team: Incorporate publicly available company information into the research process, while retaining the source, collection date, and field definitions.

These are use cases; they do not mean that Exa will automatically assess the reliability of evidence for you. Just because a web page is retrieved does not mean that the claims made on that page are valid.

How do I get started?

1. exist Exa Console Create an account and obtain an API key. First, set a team budget or a quota for each individual key to prevent test code from accidentally using up a large number of requests.
2. Test using Playground, Python, JavaScript, or HTTP requests. For general web searches, start with "Auto" and a small number of results; avoid returning a large amount of full-text content right from the start.
3. To clearly define the information goal in natural language, use `includeDomains` when you need to specify sources, rather than simply adding domain names randomly to the query text.
4. Review the returned titles, links, and excerpts, and manually check a sample of the sources to see if they actually answer the question. Then, use the same query to compare Lite's cost and results to see if they meet your needs.
5. When integrating the application, preserve the original URL, handle overages, rate limiting, and retries for failed requests, and track search costs and subsequent large-model costs separately.API Parameters - Budget and Billing Management

Below is an example HTTP request that queries only the official Python documentation and returns 5 results. The `EXA_API_KEY` should be provided via an environment variable on your local machine or server; do not hardcode the key into the front-end code of a web page:

curl https://api.exa.ai/search \
  -H "x-api-key: $EXA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "Python asyncio TaskGroup exception handling documentation",
    "type": "auto",
    "numResults": 5,
    "includeDomains": ["docs.python.org"],
    "contents": {"highlights": true}
  }'

This uses standard Highlights. Although Dynamic Highlights were officially announced in the 3.0 release notes, the current integration guide still provides examples of the dynamic mode using the Exa-Beta request header. When upgrading existing projects, you should configure them according to the guidelines for your SDK version and avoid treating `highlights: true` and `dynamic: true` as the same setting.Dynamic Mode Connection Instructions

Pricing, Free Quotas, and Concurrency Limits

For example, without using any additional billing options and with no more than 10 results per query, the search fee for 1,000 Auto retrievals is $7; this does not include the inference fees incurred when you pass the results to other large models.Exa Pricing Page

The free plan does not require a payment method to be linked. The default team-level limit for regular searches is 10 requests per second; having multiple API keys does not increase the team’s total limit. The pay-as-you-go plan can be scaled up to 25 QPS; specific terms are subject to the account and documentation. A 402 error will be returned if the quota is exhausted or a single API key’s budget is depleted; if the rate limit is exceeded, you must reduce concurrency.Free Allowances, Budgets, and Rate Explanations

Platforms, Regions, and Usage Restrictions

• Platform: Cloud-based API that can be called from HTTPS-enabled servers and supports integration via Python, JavaScript, and other languages. You do not need to purchase a local GPU to call Exa, but you do need an internet connection and a valid account.
• Lite Freshness: Returns only cached content; does not perform real-time scraping and does not accept any advanced parameters. Its low price does not mean it is suitable for tasks that require the “current state of the page.”Search Mode Restrictions
• Number of Results Returned: The maximum number of results for the public API is 100; specific modes may have different limits. Enterprise-customized capacity cannot be treated as the default quota for standard accounts.Explanation of the Number of Results
• Regional Requirements: The Terms of Service are administered by the U.S. and include requirements regarding exports and restricted entities. API connectivity, payment support, and required data residency conditions for networks in mainland China are not guaranteed based on this information and should be verified in the actual deployment environment.Terms of Service
• Data Boundaries: Queries are sent to Exa. Do not assume that standard accounts have zero data retention by default; the company lists options such as "Zero Data Retention" as features available only to eligible enterprise customers.Safety and Retention Instructions

Product Identification:The icons on this page are taken from Exa’s official brand materials (https://exa.ai/brand). The original proportions and color schemes have been preserved; they are used solely for product identification and do not imply any partnership or endorsement.

How do I choose among similar services?

The content included in a “per request” for different services is not exactly the same. When making a selection, it makes more sense to group together the same set of business problems, the same number of results, and similar output requirements, and then compare source hit rates, data freshness, latency, and total link costs—rather than simply looking at the lowest unit price.Brave Price - Tavily Credit Limit Rules - Parallel Price

How can you determine whether it's worth integrating?

Start by selecting several dozen real-world questions and check three things for each: whether the required authoritative sources can be found, whether the excerpts retain key conditions, and whether the final answers can be traced back to the original text. Then test Chinese-language questions, recent changes, and easily confused entity names.

The value of Exa 3.0 lies in its ability to provide more usable search input for AI applications. For projects that already have models but lack a stable external information source, it is an infrastructure worth trying; however, if you’re only looking up a piece of information occasionally, using existing search tools directly usually requires less setup effort.