AI for relevant search

We’re introducing a new AI-powered Brave Search technology called the Summarizer, which provides concise, concise, and meaningful answers at the top of Brave Search results based on a user’s query based solely on web search results. Unlike strictly generative AI models, which often make unsubstantiated claims, we have trained our large language models to handle multiple sources of information on the web. This allows them to produce a much more concise and precise answer in coherent language.

Moreover, the Summarizer always provides links to where the data came from. This shows the authorship of the information and helps users assess the credibility and credibility of the source, which is necessary to combat potential distortions from the authority effect inherent in large language models.

Summarizer Brave is now available to all Brave Search users on desktop and mobile. If you would like to disable the Summarizer, this can easily be done in the settings.

Using results from the web allows the Summarizer to provide up-to-date information in real time. It is important to remember that, despite the impressive results of AI models, users should not blindly believe everything that artificial intelligence produces, in the same way that you should not trust everything that is written on the Internet at all: we must always apply critical thinking to all information. that we consume.

In addition to the actual summary, our AI models can replace query-based snippets with their summarized versions, highlighting the response where possible. This allows us to provide both a main summary based on aggregation of many sources to create a comprehensive answer, and a summary of a single source (eg press articles). We provide both types of summaries together so that users can see both the general overview at the top and the snippets with highlighted answers.

It was - it became. Pay attention to the highlighted answers in the snippets. — *It was – it became. Pay attention to the highlighted answers in the snippets.*

“Brave Search is the fastest growing search engine since Bing, with 22 million queries per day providing independent search results based on our own network index. With the introduction of the AI-powered Summarizer, we are taking the next step to improve the relevancy of search results. Unlike AI chatbots, which can provide fabricated responses, the Summarizer provides a well-written summary at the top of the SERPs, aggregating the most recent sources on the web and indicating their origin for transparency. This open system is already available to all Brave Search users to better navigate search results,” says Josep Puyol, director of Brave Search.

Unlike many companies that provide similar features, we do not rely on third-party technologies and are not limited by scaling issues. The Summarizer is based on models created and managed by Brave that are tuned for maximum result generation time efficiency. To date, Brave Search is processing daily peaks of 600 requests per second, which we then run through our AI model. So far, we generate resumes for 17% of requests, and in the near future we will increase this number by scaling our system. We believe that our AI model is the largest such system to date, as it gets more traffic in requests per second than the others (we apply the Summarizer to all requests), and neither Bing nor Google did not open their systems.

In addition to scalability, we invest in the quality of generated resumes. However, since our model is still in the early stages of development, sometimes you may see “chimeras” that mix unrelated snippets into one result, or offensive or false text, but this will constantly decrease as how we will improve our models and get feedback from users.

Summarizer is fully developed by the Brave Search team and is based on our unwavering principles of independence and privacy. We do not use ChatGPT or its backend. Our AI is built on three different large language models that solve different problems, and the basic models themselves are built on open source BART or DeBERTa hosted on Hugging facewhich were largely overtrained on our own search results data.

The first model answers the question, trying to get a specific answer from text snippets, if possible. Brave already uses large language models to improve the relevancy of search results, and this model is an evolution of existing technology serving our knowledge graph and snippet capabilities. The difference lies in the number and length of snippets analyzed.
After the first response phrase, the received candidates are classified using a set of zero-shot classifiers according to a large number of criteria (hate speech, vulgarity, spam, etc.).
Finally, the final set of candidate texts is processed by a summary/paraphrasing model that tries to rewrite the input text, removing repetitions and creating a uniform text for readability.

We plan to share more technical details, with a focus on scalability, several weeks after the full release.

Please note that the Summarizer is currently disabled in Brave Search Glasses (our innovative feature that allows users to create their own filters to change the ranking of search results). We will continue to improve our models for quality compatibility with custom Glasses, and will share more information about our progress in the near future.

Conclusion

Brave Search’s Summarizer is our next step in improving search relevancy and a response to the latest changes that were driven by the release of ChatGPT last December. The news that Microsoft is going to deeply integrate OpenAI models into their Bing search engine has had a big impact on the market: despite the fact that this system is not yet publicly available, the initial feedback speaks about this model and how impressiveand even how frightening.

There is no doubt that the industry is generating a lot of hype around AI right now, but we at Brave are not yet fully convinced that large language models will radically transform search. However, when applied correctly, these models can help users with search results, which is our approach to the Summarizer. Chat-like interfaces are new technology at the moment and we are not convinced that they will be suitable for all search tasks.

However, we believe in the potential of large language models and will continue to experiment with their use not only in search, but also in the Brave Browser, where we expect truly revolutionary changes thanks to the assistant capabilities of large language models.

Brave search is available on all browser versions Brave (desktop, Android, iOS) and also available in other browsers at search.brave.com.