New AI Trick Lets Chatbots Skim Documents Like a Human
Your AI assistant could get faster, cheaper and more accurate — by reading less.
AI assistants are increasingly asked to look things up before they answer you — a technique called RAG (letting AI find facts before it speaks). But real sources are messy: a report might mix paragraphs, tables, charts and video clips. Today's systems usually grab entire chunks, dragging in irrelevant material that makes answers slower, pricier and sometimes muddier. A team of researchers — Hyojeong Yun, Jueun Kim and Wook-Shin Han — has built a fix called CANOPY.
CANOPY treats every retrieved item like an outline. A document becomes chapters, sections and sentences; an image becomes regions. The system then scores each region against your question and keeps only what actually helps, zooming in or out as needed. If the evidence still looks thin, a built-in "critic" sends the system back to search for more, compressing anything new before adding it. Crucially, most of this trimming happens without calling a large AI model, which keeps it fast and cheap.
The results are promising. Across five question-answering tests over a 33-million-item collection mixing text, tables, images and video, CANOPY beat the comparison systems on average accuracy. When paired with an AI model called Qwen3-VL-8B-Instruct, it shrank the amount of text the answering model had to read by 14–28% while keeping accuracy about the same. The biggest gains came on "multi-hop" questions — ones that require stitching together several facts — mostly because the system went back for extra searching.
The honest catch: this is a research paper, not a product you can download. The tests were run in a lab, on a specific pile of data, and gains depend on having a good search index underneath. Still, the direction is clear and it matters. AI that reads less and answers better means lower bills for companies, faster replies for you, and less of your data being fed into a model just to get a simple answer.
- CANOPY helps AI search tools throw away irrelevant material before answering, instead of swallowing whole documents.
- In tests over 33 million mixed files, it cut the text the AI had to read by up to 28% with no loss in answer quality.
- It's still academic research — no app yet — but it points toward cheaper, faster, more private AI assistants.
Why It Matters
Cheaper, quicker AI answers — and less of your data fed into models just to get a simple reply.