LiteRAG: The AI Lookup Trick That Cuts Costs by 99%

Every time you ask an AI chatbot a tricky question, it often looks things up first. The technique is called RAG (letting AI search before it answers). The fanciest version, GraphRAG, builds a map of how facts connect, which helps with questions that need several pieces of information joined together. The problem: an AI model has to steer the search every step of the way, so answers get slow and expensive. A team of researchers in Spain has now published LiteRAG, a leaner version.

Instead of hiring an AI to make decisions mid-search, LiteRAG uses plain math and simple rules. Think of following signposts on a road map instead of paying a tour guide at every corner. On DistComp, a test set of academic papers about computer systems, LiteRAG scored the highest quality of all methods tested (0.798) while running more than 100 times faster and costing over 99% less than GraphRAG Global and DRIFT. On another test set, UltraDomain, it matched a rival called LinearRAG while using about 14 times fewer tokens — tokens are the word-chunks AI companies bill you for.

Why should you care? Because speed and cost savings eventually reach you. If AI companies pay a fraction of the price per question, they can charge less, answer instantly, and handle longer documents without hitting limits. Faster searching also means less waiting and fewer half-right answers, especially for questions that require connecting several facts — like 'which of my contracts renew this year, and which have penalty clauses?'

The catch: this is a research paper, not a product you can sign up for today. It was tested only on specialist academic datasets, and it still needs a well-organized map of your documents built in advance. Messy real-world data and commercial versions may behave differently. Still, it points in a clear direction: AI's cost of thinking keeps falling.