Exa, the AI-native search engine, has unveiled a significant update to its Exa Deep endpoint. This revamped system acts as an agent designed to tackle complex search tasks, promising higher quality results through advanced query expansion and LLM reasoning for search.
The updated Exa Deep is engineered to be faster and more cost-effective than its predecessor. A key advancement is its ability to deliver structured JSON outputs, complete with field-level citations that ground the information directly to its sources. This makes it particularly valuable for detailed research across various domains.
An Agent for Every Search
Exa Deep's architecture now utilizes the company's Exa Instant search alongside LLM reasoning. This combination allows it to better understand user intent, generate multiple search agents concurrently, and synthesize findings with verifiable citations. This approach positions Exa Deep as a foundational primitive for building sophisticated research agents, potentially replacing complex orchestration layers with a single API call.
The company claims Exa Deep outperforms competitors like Perplexity's Sonar Reasoning Pro and Parallel Task API in benchmark tests, demonstrating superior speed and accuracy. The web is increasingly being rebuilt for AI search agents, and Exa Deep aims to be a critical component in this evolution.
Structured Outputs and Grounding
Users can now define an output schema, enabling Exa Deep to return precisely structured data. For example, a query for aerospace companies and their CEOs could yield a JSON object containing company names and CEO details, with each piece of data meticulously linked back to its original source via citations and confidence scores.
This structured output capability is crucial for applications requiring precise data extraction and verification, such as financial analysis of SEC filings, surveying scientific literature, or ongoing news monitoring. The platform's ability to synthesize information from multiple documents and provide grounded citations addresses the growing need for reliable, AI-driven research tools, akin to how Neural RAG Redefines Web Search for the AI Era.
Pricing and Performance
Exa Deep offers two tiers: 'deep' with latency between 4-12 seconds at $12 per 1,000 requests, and 'deep-reasoning' with 12-50 second latency at $15 per 1,000 requests. The standard 'deep' tier is now 20% cheaper.
Exa suggests that the future of search lies in agents that can decompose complex queries, search in parallel, and iterate to find answers. As AIs increasingly perform web searches, the need for search engines that can reason rather than merely retrieve data becomes paramount. Exa Deep represents the company's answer to this evolving landscape.
