Frequently asked questions

What's different about Exa Search?

Exa search is fully neural. We use a transformer-based model to understand your query and return the most relevant links. Exa has embedded large portions of the web so you can make extremely specific and complex queries, and get only the highest quality results.

How is this different from Google?

Google search is mostly keyword based, meaning that it matches the words of your query to the words in links. For example, a Google search for “companies working on AI for finance” returns almost all links like “Top 10 companies developing AI for financial services”. In contrast, Exa’s neural search understands meaning and therefore returns actual company urls.

How can Exa be used in an LLM?

Exa supercharges LLMs with high quality, relevant web content to avoid hallucination and outdated text generations. An LLM can take a user’s question, use Exa to find relevant web content, and answer the question based on reliable information.

How does Exa compare to other search APIs?

Exa is a brand-new search engine, built for use by LLMs.

  • We handle natural language queries.
  • We can instantly return page content for any page in our index.
  • We can return thousands of results for automatic processing.

How often is the index updated?

We update our index every two minutes, and are constantly adding batches of new links. We target the highest quality web pages. If there’s a specific website you need indexed, please reach out!

Is it free?

Exa is free to use up to 1000 requests per month. Past that, look at our pricing page.

What's our roadmap?

  • Custom comprehensive datasets
  • Support arbitrary non-neural filters.
  • Train a much bigger version of our model
  • Solve search. No, really.

How does similarity search work?

When you search using a url, Exa crawls the url, parses the main content from the html, and searches with that parsed content.

The model chooses webpages which it predicts are talked about in similar ways to the prompt url. That means the model considers a range of factors about the page, including the text style, the domain, and the main ideas inside the text.

Similarity search is natural extension for a neural search engine like Exa, and something that's difficult with keyword search engines like google.

Does the search use ChatGPT under the hood?

Nope, we train our own search model from scratch! You can't use a language model like ChatGPT to do do straight search because it would have to memorize the internet, which is not possible from an information theory perspective.

Why do searches need to be prompt engineered?

Exa uses a transformer architecture to predict links given text, and it gets its power from having been trained on the way that people talk about links on the Internet. This training produces a model that returns links that are both high in relevance and quality.

However, it also means the model expects queries that look like how people describe a link on the Internet. For example, 'best restaurants in SF" is a bad prompt, whereas "Here is the best restaurant in SF:" is a good prompt.