Exa Gives AI Search a Time Machine: 400 Billion Webpage Snapshots to Prevent Agents from Peeking into the Future

Beating AI News Flash: Exa launches Snapshot, allowing searches to directly "go back in time." It has indexed over 400 billion web snapshots spanning approximately 20 years. Developers can specify a time, and Exa's search and web reading API will return the web page versions that were saved at that time.

The most direct use is to prevent AI evaluations from "peeking at answers." For example, if a question is designed in June and only given to an Agent in September, the answer may have already appeared in papers, GitHub PRs, or blogs. Now, the search time can be locked to May, so the Agent cannot see answers that were only made public later. Financial teams can also use the same method for historical backtesting, showing the model only information that already existed at the time.

However, this "time machine" currently still has limitations. It can guarantee that what you see is the web page version from that time, but the search results are not necessarily exactly the same as they were back then. Ordinary paid users can currently only look back as far as the most recent 5 months; to search earlier history, they need to contact Exa.