Point-in-time search
Point-in-time search is a search that returns results as the index stood at a chosen moment in the past: pages acquired after that moment are excluded, so the answer is the one the search engine would have given then. It is not a recency filter. To get recent pages, a search sets a lower bound on the publication date.
It is used to replay past agent runs, to backtest forecasting and trading research without look-ahead bias and to build evaluations that a model cannot pass with knowledge of later events. In the Keenable Search API, point-in-time search is the query_time parameter.
Example
A researcher checks whether an agent could have predicted the result of a vote held on November 5. With point-in-time search set to the morning of November 4, no page reporting the result can appear, so the agent’s forecast can be scored fairly.
FAQ
How is point-in-time search different from a date filter?
A publication-date filter keeps pages published within a range in today’s index. Point-in-time search rewinds the index itself, so pages that had not been crawled at that moment are excluded even if they are about older events.
How do you run a point-in-time search with Keenable?
In the Keenable Search API, set query_time to a timestamp or a date. Relative date filters in the same request, such as published_after set to 30d, then count back from query_time instead of from the request time.
Related terms
Written by Keenable.
Published
