Microsoft Technology Licensing LLC published a patent on August 13, 2026, outlining a system designed to compress and filter web sources for generative artificial intelligence models like Bing Copilot.
The patent, US20260236548A1, describes methods intended to address token limits and enhance the quality of responses from AI search engines by processing what it terms “grounding data.”
The system prioritizes sentences most connected to the rest of the text and the user’s query, according to the patent. It achieves this by employing a similarity graph and favoring content found in introductory or concluding sections of a text.
Content filtering also incorporates freshness criteria, utilizing domain-specific thresholds to ensure relevance. Additionally, a reputation score is applied, which combines historical accuracy, domain authority, and peer validation to assess source reliability.
The patent aims to mitigate common large language model (LLM) issues such as “Lost in the middle” and “Recency bias,” according to the document. It seeks to achieve this by compressing data into a single API call, streamlining the information retrieval process.
Techniques detailed in the patent include extractive summarization and clustering by semantic proximity. It also prioritizes sentences containing frequent and relevant terms, alongside a discourse-based method for content selection.
Ian Sorin is listed as an inventor on the patent. The methods described could influence future AI model performance and content creation strategies for generative AI search engines, Microsoft reported.
Source: Abondance
