Brave Browser Introduces Summarizer for Search Based on In-House LLM
Not quite conversational search, but more than just links
Everyone who is not Google expects online search features and user habits to change very soon. Google may believe this now; thus, the all-out effort to bring Bard to market. However, four months ago, the company clearly did not see this change as imminent.
It is fair to say that few companies thought the window of opportunity around search would open this quickly, but here we are. Brave, a small player in the browser market, and a smaller player in the search market, just introduced a new search Summarizer feature based on a large language model (LLM). According to an announcement this week:
The Summarizer provides concise and to-the-point answers at the top of Brave Search results pages, in response to the user’s input, solely based on Web search results. Unlike a purely generative AI model, which is prone to spout unsubstantiated assertions, we trained our large language models (LLMs) to process multiple sources of information present on the Web. This produces a more concise, accurate answer, expressed in coherent language.
In addition, the provenance of original sources of data is cited at all times via links. This maintains the rightful attribution of information, and helps users assess the trustworthiness of the sources, both of which are needed to mitigate the authority biases of large language models.
Summarizer the Answer Box
I gave Summarizer a test drive and found it to provide a consistent experience when it shows up on the page. The summarization feature does not show up with every search. This surprised me after using the New Bing, You.com, and Perplexity.ai, which all deliver the integrated summary answer every time.
With Brave, I received a Summarizer in response to less than half of my submitted questions. This was not a representative or large sample of queries, so don’t take this as a scientific finding. However, I found it noteworthy that it was often hard to get the Summarizer to appear at all. At first, I thought I had an issue with the Browser update.
Summarizer, instead, is more like a Google answer box. Google has provided short summary answers for years for a variety of queries. The short answer with a link to the source is fairly similar to what Brave is providing with Summarizer.
Google Answer Box frequency likely peaked around six years ago at around 25% - 30%, but recent studies suggest the incidence is consistently below 15%. It is not clear why Answer Boxes fell out of favor with Google, but Brave is bringing back a similar model even if the underlying technology is different.
How Summarizer Works
Brave stressed in the announcement that the feature is not powered by GPT-3 or OpenAI. It is based on internally developed LLMs and relies on three separate models.
Question Answering: This model attempts to extract an answer from text snippets Brave is already extracting from crawled websites.
Filtering: Snippet results “are further classified with an ensemble of zero-shot classifiers on a wide variety of criteria (hate-speech, vulgar writing, spam, etc).”
Summarizer/paraphraser: The best answers are summarized into a single, coherent response.
The model is generating responses from known sources, so it is fairly easy to provide a link back to the web pages that served as sources for the response. With Bing or Perplexity, you will often see 4-6 source links. Brave Summaizer appears to typically default to a single source — similar again to a Google Answer Box — or sometimes two sources. There may be instances where more sources are provided in the Summarizer box, but I did not observe any.
Below the Summarizer box are links to articles you would expect from a traditional search results page. Summarizer is simply a top-of-page augmentation to that search experience.
Not Conversational Search
Bing and Perplexity search are also different in their conversational search features. After you receive a summary answer, you can “chat” with the search to refine or expand on your questions. Summarizer is a single-shot response. There is no context window or conversational interaction model. It is all fire and forget.
So, when it comes to search, Summarizer is among the least interesting of the new models we have evaluated. However, it remains noteworthy because it represents another example of how search features are changing, and it is a good application of LLMs that balances the benefits of more user-friendly search results with the need to avoid producing incorrect information.
