Microsoft Unveils New Deep Search Feature Leveraging Generative AI Follow Dec 05, 2023 · 1 min read
Microsoft Unveils New Deep Search Feature Leveraging Generative AI
Share this

Microsoft Corp. today announced the launch of Deep Search, a new optional feature for Bing powered by generative AI. Deep Search helps searchers find comprehensive answers to complex questions that may not have simple keyword-based responses.

How It Works

Built atop Bing’s existing index and rankings, Deep Search utilizes the GPT-4 model to discern all possible intents behind an ambiguous query. It then generates descriptions for each potential meaning in order to return an ideal, wide-ranging set of results.

After querying the index in multiple ways, Deep Search surfaces information that would not typically appear high in regular search listings. For example, when a user searches “how do points systems work in Japan,” Deep Search may determine other useful search phrases like “comparison of loyalty programs by category Japan” and “managing loyalty points with phone apps.”

As Microsoft stated, “By doing this, Deep Search can find results that cover different aspects of my query, even if they don’t explicitly include the original keywords.” They noted that while Bing currently evaluates millions of pages per search, Deep Search examines 10 times more content to deliver more informative and targeted findings.

Key Features and Availability

An animated demo shared by Microsoft illustrates Deep Search in action. Pages rank highly if they strongly match Bing’s expanded query description. Other key relevance factors include topic match, level of detail, source credibility, freshness and popularity.

One current downside is speed, as results can take up to 30 seconds to load. The feature is also only available to small, random test groups globally for now. Microsoft indicated they created Deep Search to give “more relevant and comprehensive answers to the most complex search queries.” They noted it is meant to enhance, not replace, Bing’s existing functionality.

Written by Follow