Skip to main content
Article
100%
HD

Have you ever tried searching for something online and felt frustrated because the results had nothing to do with what you actually had in mind? It happens because, sometimes, the way we phrase our questions isn't exactly what computers expect. But here is the good news: researchers have just invented a new trick to make artificial intelligence much smarter at understanding our desires!

It is called HiPC-QR, and it uses a technique known as 'prompt chaining.' Imagine having an assistant who, instead of rushing straight to find the answer, takes a moment to re-read your question. First, it extracts the most important keywords and then filters out all the extra fluff that might cause confusion. It is like if, before heading to the library, an assistant cleaned up your request by removing useless words and adding the right ones to find exactly the book you need.

This method does more than just look for synonyms, like older systems used to do. Thanks to Large Language Models, AI can now understand the intent behind your words. Studies have shown that this system is much more accurate at finding the right results compared to previous methods. Essentially, we are moving from 'trial and error' searching to 'precision' searching, making our digital lives much smoother and faster!

Bonus
100%
HD

The central problem addressed by the HiPC-QR research is the 'isalignment' between user intent and the AI-reformatted query. Often, when we ask a language model to improve a search, the AI tends to generate overly long sentences or include unrequested concepts, which act as 'noise' during the document retrieval phase (Information Retrieval).

The HiPC-QR approach solves this problem through a two-step hierarchical structure. In the first phase, the model focuses on semantic extraction: it identifies the core concepts of the original query. In the second phase, structural refinement occurs: the model evaluates which keywords are essential to the user's intent and which are redundant or misleading. This 'filtering and expansion' process allows for a balance between two often conflicting goals: precision (finding only what is needed) and recall (finding everything that is relevant).

Tests conducted on industry-standard datasets, such as MS MARCO and TREC Deep Learning, confirmed that this method significantly improves crucial metrics like Recall@10 (the ability to find the correct document among the top ten results) and MRR@10 (Mean Reciprocal Rank, which measures how high in the list the relevant result appears). In short, using structured prompt chains allows for leveraging the vast knowledge of Large Language Models without the risk of generating queries that are either too generic or overly specific.

Ask AI

Explore key aspects of this story with 3 quick insights

Source

Authors:
Hua YangHanyang LiTeresa Gonçalves
Date Published:
Original Article www.mdpi.com
Fact-check verified on 01/08/2026
Read our Editorial Guidelines & AI Ethics Policy
Creative Commons License CC BY-NC-SA 4.0

Keep exploring

Discover the full archive of science bites or head back to home

Riconnessione al server... Aggiornamento della sessione in corso
Connessione interrotta Impossibile stabilire il collegamento
Sessione aggiornata Ricarica per visualizzare i contenuti