New AI File Shortcut Makes Search Engines Up to 8x Faster

When you type a question into a search engine, it first pulls up a list of possible matches, then runs a second step to sort those matches by usefulness. This second step is increasingly done by AI models that read the full text of each result. But if millions of people ask different questions, the same document gets read over and over again, wasting time and computing power.

A team of researchers has come up with a clever workaround called DoPR. Their idea: pre-process each document once, squeeze its meaning into a compact digital summary, and store that summary. When someone asks a question, the AI only looks at the question and the summary instead of the whole text. It is like reading a book's back cover instead of the entire book every time.

In tests using standard search quality benchmarks, this method made the sorting step up to 8 times faster and used 8 times less memory. Best of all, the quality of the results barely dropped, staying between 97.1% and 99.5% of what full-reading models achieve. The researchers also showed it works across different AI model sizes and search tasks.

The practical upside: faster, cheaper search systems. That could mean smoother experiences on your favorite search engine, lower data center energy use, and more affordable AI features in apps. The catch is that the document summaries must be built ahead of time, so new or frequently updated documents need a refresh. Still, for searches over stable collections of documents - like legal records, medical libraries, or company files - this could be a big win.