What Is Tavily? A Search API for LLMs and Its Alternatives
Tavily is a search API built for LLMs and RAG. See what it returns, how it differs from Brave, Exa and SERP APIs, and how to add search with no extra account.
Tavily is a web search API designed for language models and agents. Instead of returning ten blue links for a person to read, it returns cleaned, ranked snippets and, optionally, extracted page content that you can put straight into a prompt. It is popular in RAG pipelines and agent frameworks such as LangChain and LlamaIndex.
What it does
- Search: takes a query and returns results with titles, URLs and relevant content excerpts, plus an optional short generated answer.
- Depth modes: a basic mode for quick lookups and an advanced mode that reads more of each page and costs more per call.
- Extract and crawl: fetch and clean the text of specific URLs, or walk a site.
- Filters: topic (general or news), time range and domain include/exclude lists.
Recent change: Tavily was acquired
On 2026-02-10 Nebius, an AI cloud company, announced it was acquiring Tavily. Public pricing had not changed at the time of writing. If you build a product on any single search vendor, an acquisition is a reason to keep your integration easy to swap.
How it compares
| Type | Examples | What you get |
|---|---|---|
| AI-oriented search API | Tavily, Exa, Parallel, Perplexity Search API | Ranked results plus text meant for an LLM prompt |
| Independent index | Brave Search API | Its own web index, raw ranked results |
| SERP API | SerpAPI, Serper | A structured copy of a Google results page |
| Model-native search | OpenAI, Anthropic, Google built-in tools | Search inside one provider's models only |
Rules of thumb: choose an AI-oriented API when the results go into a prompt; a SERP API when you need exactly what Google shows (rankings, ads, maps); an independent index when you want a non-Google source. Model-native tools are convenient but tie you to one model family.
Wiring it up yourself
A typical setup has four steps: create an account and key with the search vendor, call the search endpoint, trim the results to fit your context window, and inject them into the prompt with a request to cite sources. You also handle rate limits, failures and a second bill.
The alternative: search as a request option
If you already call models through BazaarLink, you can skip that plumbing. Add :online to the model name, or a web plugin option, and the request is searched first and answered with citations:
{
"model": "openai/gpt-6-luna:online",
"messages": [{"role": "user", "content": "Summarize this week's news on open-weight models"}]
}
You can set depth (low, medium, high, fast), topic, time range, country and language in the same request. Searches are billed per search, and a search that fails is not charged. See our guide to adding web search with the :online suffix.
Which should you choose?
If you need crawling or a dedicated extraction pipeline, use a dedicated search vendor directly. If you want your existing model calls to gain live, cited answers with the least code and a single bill, a request option is simpler. Whichever you pick, keep the search step behind a small function so it can be replaced.
For the four tools and pricing, see Tavily-style AI search API: Search, Extract, Crawl & Map.
TWD billing · Taiwan invoices · leading AI models · OpenAI-compatible API