Exa vs. Tavily

Exa scores 10 points higher on complex retrieval in independent enterprise testing.

Most search APIs handle simple factual lookups. The queries that break things are multi-hop: 'What was the revenue impact of this acquisition?' requires finding the announcement, then the earnings report, then extracting the figure. If the search API can't follow that chain, the agent silently fails.

Tavily's keyword-based retrieval keeps up on simple lookups but falls behind when queries need semantic understanding or multi-step reasoning. The harder the query, the wider the gap.

p95 latency under 1.7 seconds, with an instant mode under 200ms.

An agent that chains 50 searches per workflow loses minutes to a slow API. For voice and real-time AI, anything over 500ms breaks the interaction. In the Fortune 100 enterprise evaluation, Exa's p95 ranged from 1.4s to 1.7s across three benchmarks. Tavily's ranged from 3.8s to 4.5s.

Tavily reports 90ms for its ultra-fast mode, but that number only holds for the simplest queries. On typical evaluation queries: 210ms. On longer queries: 420ms. Exa Instant delivers sub-200ms with no query-length penalty.

Search 1B+ people and companies by role, location, and industry.

Sales enrichment, recruiting agents, competitive intelligence, and due diligence all require searching for specific people and companies, not web pages. General web search returns LinkedIn profile pages if you're lucky, but can't filter by role, funding stage, or employee count. Tavily has no people or company search.

Exa indexes over 1 billion LinkedIn profiles with 50M+ updates per week, plus a dedicated company search index built on proprietary embeddings. Queries like 'Series B fintech companies in Singapore with 50-200 employees' or 'VP of Engineering at mid-market healthcare companies' return structured, current results. The company index covers entities that general web search misses: pre-Series A startups, niche B2B verticals, regional players with minimal web presence. Both benchmarks (~800 company queries, 1,400 people queries) are open-sourced on GitHub.

Code search extracts the relevant snippet, not the whole docs page.

Coding agents that pull full documentation pages into their context window waste tokens on navigation, sidebars, and unrelated sections. The agent needs the relevant function signature or usage example, not the surrounding page chrome. Exa Code searches 1B+ web pages and returns token-succinct excerpts, typically a few hundred tokens per result. The maxCharacters parameter lets agents set exact context budgets: 500 characters for a quick API reference, 5,000 for a full tutorial. Tavily has no code search.

Query-dependent highlights cut tokens 50-75% while improving RAG accuracy.

RAG pipelines that send full page text to the LLM pay for irrelevant tokens and get worse answers. Exa's highlights extract the passages that actually answer the query, improving RAG accuracy by 10% in internal testing at 50-75% fewer tokens. The maxCharacters parameter sets a per-page token budget so agents control exactly how much context goes downstream.

Tavily returns static page content via include_raw_content, the same text regardless of what was searched. Searching for 'React Server Components' against a Next.js docs page returns the same content as searching for 'deployment configuration' against that same page. All filtering happens in your LLM, at your token cost.

1,200-domain filters and text matching that hold under production load.

Production deployments need to restrict results to approved sources. An IT support agent that returns results outside its domain allowlist is a security issue, not a quality issue. Exa's includeDomains and excludeDomains accept up to 1,200 entries each. Tavily caps at 300 include and 150 exclude.

Exa also has includeText / excludeText for content keyword matching, ISO 8601 date ranges on both publish and crawl dates, category filters (company, people, news, tweet, research_paper), livecrawl control via maxAgeHours, and language filtering. Tavily has date presets, country boosting, and safe search, but no category filters beyond news and finance.