Exa Search API - Real-Time Web Data for AI Agents

TL;DR

TL;DR: Exa is a search API specifically designed for AI agents, offering live web scraping, deep research, and structured content retrieval from a single endpoint — with a claimed 20x recall advantage over traditional search engines.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: exa.ai — verified
• Documentation: docs.exa.ai — verified (Mintlify)
• HN launch thread: news.ycombinator.com/item?id=43175781 — YC S21, 28 points
• YC Companies page: ycombinator.com/companies/exa — accessible, YC S21 confirmed

What Is Exa?

Exa positions itself as a search engine built from the ground up for AI agents, not for humans. While traditional search engines optimize for click-through rates and human readability, Exa indexes content in a way that makes it directly machine-consumable — structured, dense, and retrievable via semantic search.

The core product is a REST API that accepts a query and returns structured web content. According to the product page, the index covers billions of documents crawled daily, with vector embeddings computed for every page. This allows Exa to do semantic similarity search across the live web, returning results that match intent rather than just keyword overlap.

The product claim is striking: “>20x recall than Google” for the same query, as stated in the Show HN launch. Whether that claim holds in production depends heavily on query type and evaluation methodology — worth treating as a directional claim rather than a benchmark.

Core Features

The Exa API covers three main retrieval scenarios:

Live Search — queries the live web in real time. Returns URLs, titles, summaries, and raw HTML snippets. Suitable for news, pricing pages, event data, or anything that changes frequently.

Deep Research — returns fully scraped and parsed content from the top results. Rather than returning links, Exa returns the actual content, structured and deduplicated. The API handles extraction, deduplication, and citation.

Semantic Search — uses vector embeddings to find conceptually related content. This is the mode that differs most from keyword search: a query like “what did companies say about AI safety in Q1 earnings calls” returns relevant pages even if no exact phrase matches.

Additional capabilities visible on the product page include:

• Full-text content extraction
• Page-level and site-level filtering
• Date range filtering
• Category and domain whitelisting/blacklisting
• Streaming response support

Setup Workflow

Step 1: Get an API Key

Sign up at exa.ai to receive an API key. The free tier includes a daily request quota suitable for local development and experimentation.

Step 2: Install the SDK

Exa provides official SDKs for Python and JavaScript/TypeScript.

pip install exa-py

npm install exa-search

Step 3: Make Your First Request

from exa_py import Exa

exa = Exa(api_key="YOUR_API_KEY")

# Live search
result = exa.search(
    "Show HN launches from YC S21 companies in 2026",
    num_results=10,
    text=True
)

for item in result.results:
    print(item.title, item.url)

import Exa from 'exa-search';

const exa = new Exa(process.env.EXA_API_KEY);

const result = await exa.search({
  query: "Show HN launches from YC S21 companies in 2026",
  numResults: 10,
  text: true
});

result.results.forEach(item => {
  console.log(item.title, item.url);
});

Step 4: Deep Research Mode

# Get full content from top results
research = exa.search(
    "LLM context window developments 2026",
    num_results=5,
    text=True,
    highlights=True,
    subpages=True  # follow links within pages
)

for item in research.results:
    print(item.highlights[0] if item.highlights else item.text[:200])

How Exa Differs from Google Custom Search or SerpAPI

Traditional search APIs like Google Custom Search or SerpAPI return a ranked list of URLs and snippets. The agent still needs to fetch, parse, and extract information from each URL — a multi-step process that adds latency, failure points, and token cost.

Exa collapses this into a single call by pre-indexing the web with embeddings and running the retrieval + extraction pipeline server-side. The trade-off is that Exa’s index has a refresh lag (how fresh “live” actually is varies by query and crawl frequency), whereas Google Custom Search queries the index in real time.

For agents that need to read many pages to answer a question, Exa’s deep research mode reduces the round-trip count significantly. For agents that need guaranteed freshest data (e.g., real-time stock prices), a live search API is still necessary.

Practical Evaluation Checklist

• [ ] API key obtained and SDK installed
• [ ] Basic keyword query returns expected results
• [ ] Semantic query returns conceptually related results that keyword search misses
• [ ] Deep research mode returns extracted content without needing to fetch URLs separately
• [ ] Streaming response works for large result sets
• [ ] Date filtering correctly constrains results to specified time ranges
• [ ] Domain filtering works for both whitelisting and blacklisting
• [ ] Rate limits understood for production use

Security Notes

• API keys grant read access to your Exa quota. Treat them like passwords — never commit them to version control or expose them client-side.
• Exa runs on Cloudflare’s network. Data handling and retention policies are governed by their privacy policy.
• For production deployments, consider setting up key-scoped rate limits via the Exa dashboard.

FAQ

Q: Does Exa replace Google Custom Search? A: It depends on your use case. Exa is stronger for semantic and conceptual retrieval, and it pre-extracts content server-side. Google Custom Search offers real-time index access and is better for queries where freshness is critical. Many production systems use both.

Q: What does “20x recall” mean in practice? A: The claim from the Show HN launch refers to a specific benchmark comparing embedding-based semantic retrieval against Google’s keyword matching on a set of complex queries. Your mileage will vary — simple factual queries often perform similarly across engines.

Q: Is there a free tier? A: Yes, Exa offers a free tier with daily request limits. Volume pricing for production use is available on the pricing page at exa.ai.

Q: What programming languages are supported? A: Official SDKs exist for Python (exa-py) and JavaScript/TypeScript (exa-search). The underlying REST API can be called from any language with HTTP support.

Conclusion

Exa fills a specific gap for AI agents: semantic, pre-extracted web content in a single API call. The embedding-based index gives it a different strength profile than keyword search, particularly for complex, multi-hop queries. If your agent needs to research, compare, or synthesize information from many web sources, the deep research mode alone is worth evaluating — it removes the need to build your own web scraping and extraction pipeline.

The YC S21 backing and active development (the product launched on HN in mid-2025 with ongoing updates) suggest this is a production-grade service rather than a research prototype. Worth adding to your agent toolbelt if web retrieval is part of your workflow.