Exa

Exa provides web search and page-content APIs for applications and AI agents. Its Contents documentation describes extracting text, relevant highlights or summaries from supplied URLs, including JavaScript-rendered pages and PDFs. The right output depends on whether your task needs a short source passage, a complete document or a generated overview.

Keep extracted evidence separate from generated summaries

Exa's highlights return passages copied from a page. Full-text output returns the cleaned page body as Markdown. These can preserve source material for an agent to inspect, although an excerpt may leave out context that changes its meaning. Keep the page URL beside important claims.

The summary option makes a language-model call for each page and can return an overview or fields requested through a JSON schema. Treat that generated output as an interpretation rather than a quotation. For factual work, compare it with the underlying text before storing it as evidence or publishing a claim.

Choose freshness and bounds for the request

The documentation explains that maxAgeHours controls whether cached content is acceptable. A value of zero requests a fresh fetch, while minus one uses cached content only. A positive value allows content newer than the specified age. Choose the policy according to how fast the information changes.

Exa also exposes a timeout for fresh crawling and limits for extracted text. Check each URL's returned status so a failed fetch does not appear to be a page with no useful content. Bound the amount of context you pass downstream, but keep enough of a source to understand a qualification or limitation. A fast cached response is not proof that a price or availability statement is current.

Account for the requested content views

The guide recommends one content view per request and says requesting highlights, text and summary together returns and bills each separately. It also describes including content with Search results, with a documented allowance before additional page charges. Check the current API pricing and endpoint schema before estimating a recurring workload.

For a source-grounded pipeline, store the actual publisher URL and distinguish retrieved text from generated fields. Confirm the applicable handling of supplied URLs and request data before processing sensitive material. Search and extraction can locate evidence; they do not establish that a publisher's statements are true or that you have permission to reproduce its full text.