Indian startup makes a Google competitor, co-founder Vinci Mathews explains why Nofrills AI is a hybrid of a search engine and an AI chatbot

For example, consider Wikipedia, which relies on a database to deliver results. Our engine, on the other hand, combines a database with live search capabilities. It goes out to the web, filters results, and uses our proprietary algorithm to understand your query. We then process these results to provide you with the most accurate information. Additionally, we show you the sources from which we derive the results. It’s a multifaceted approach to search. But I must say we’re not even competing with Google and our solution is completely different from what Google is offering right now. It’s not going to be easy but there’s going to be learning. The more people use it, the more it becomes intuitive.

Is searching the web still the same, or has it changed?

NoFrills AI is part search engine and part AI chatbot. (Image credit: Nofrills)

Vinci Mathews: I would say search has evolved significantly over the years. Traditional search engines such as Google and Bing primarily rely on keyword matching and paid optimisation to deliver search results, often burying the most relevant information deep within pages of links. This can make it difficult for users to find accurate and reliable information quickly. These engines struggle with natural language queries and follow-up questions, treating each as a separate search. For example, if you search ‘Hotels near me’ on Google, you might get a long list of links. However, if you want to know, ‘How many offer free flexible checkout time?’ you’d have to start a new search, often ending up with the same irrelevant links due to the limitations of keyword-based indexing. Nofrills AI is designed to bridge this gap by understanding natural language, providing a concise summary of the information users seek, and directing them to the exact sources (URLs) containing detailed content. The short summary from Nofrills’ sources helps users avoid unnecessary detours and land on the exact information they’re looking for. You can also ask unlimited follow-up questions in a thread, something traditional search engines lack.

With Nofrills AI, what is the opportunity here to make something that Google simply couldn’t or wouldn’t do?

Vinci Mathews: The key distinction between Google and other traditional search engines lies in its ranking and indexing algorithms. While these algorithms have made Google the dominant force in traditional search, they also limit its ability to evolve into a reliable AI-driven search engine. Google’s algorithms are primarily optimised for search engine optimisation (SEO) tactics, making them highly effective for standard keyword searches. This reliance on SEO, traffic, and backlinks ensures that Google’s rankings are usually accurate for such searches. However, problems arise when users ask questions in natural language. In these cases, Google’s rankings often fail to surface sources that contain the correct answers, which are frequently buried on the second or third page. This limitation is evident in Google’s attempts to incorporate AI summaries into search results, which have been criticised for occasionally providing dangerous suggestions. In contrast, Nofrills AI has trained a smaller model with enhanced reading comprehension, which is highly effective and accurate in interpreting summaries and answers from a wide range of URLs, regardless of their ranking.

Nofrills AI’s ability to handle large volumes of sources without being restricted by context limits allows it to overcome the biases and clickbait tactics that often influence traditional ranking systems. While Google may eventually introduce a new ranking system tailored for AI-based search or overhaul its current system for this purpose, this transition will take time. This delay creates a valuable opportunity for us to pursue.

Is Nofrills AI trying to be the best of both worlds: a search engine and an AI chatbot?

Vinci Mathews: You can say Nofrills AI is positioned as a hybrid of a search engine and an AI chatbot, building on the strengths of both technologies to provide a unique user experience. It operates as a conversational search engine, with real-time answers sourced from the web, while seamlessly integrating elements characteristic of chatbots, including advanced natural language processing and engaging interactive dialogue capabilities.

You mentioned that Nofrills AI is different from a Google search. Does the user interface also play a significant role in obtaining the correct answer to a query?

Vinci Mathews: The difference between Google and us is that with our product, the interface feels like you’re asking someone a question. For example, if I want to find Indo-Chinese food that specialises in vegetarian options, I can type exactly that and I will get an answer. So it’s a conversation-based search engine, not keyword-based.

Building a search engine is hard, but building a better search engine and running it is even harder, especially when more users join in. What do you have to say?

Vinci Mathews: Building a search engine is often seen as complex, but it’s more accessible than many think. You don’t need to start from scratch with a crawler; instead, you can leverage existing resources like Common Crawl (an open-source crawl repository) and an open-source index server. For fresh data, incremental spidering is sufficient. The real challenge lies in developing a ranking algorithm that’s even a fraction as effective as Google’s. This is where the real effort and potentially years, if not decades, of work come in. Nofrills takes advantage of Common Crawl, integrating RSS feeds for the latest updates and using APIs for real-time data. However, what sets us apart is the use of AI to address ranking challenges. We use suboptimal search results as context for our model, which is then able to accurately identify the most relevant answers from the available sources.

How do you rank all those pages, in order, for every single query your search engine might get?

Vinci Mathews: We use a semantic-based ranking model to generate an initial set of sources for each user query. Additionally, we’ve trained a specialised model to enhance its reading comprehension, enabling it to sift through poorly ranked search results and identify the most relevant ones. Our model is designed to thoroughly evaluate each source presented, ranking the most useful ones and providing citations to these sources. It’s akin to a human meticulously reviewing 30 URLs for a query, identifying the most relevant content, and interpreting the results. However, instead of relying on a human, our model performs this task automatically. This approach allows us to bypass the need for a highly complex ranking algorithm. Instead, our semantic-based minimum ranker, supported by an advanced model, delivers accurate results and helps us compete with the competition.

Currently, NoFrills AI can be accessed on the web, but a mobile app is already in the works. (Image credit: Nofrills)

Do you have a mechanism through which you filter out problematic content or a web page?

Vinci Mathews: To ensure that our interactions are safe and respectful, the team has implemented a robust prompt guard system. This system is designed to detect and block any questions or prompts that involve moral turpitude, unethical behaviour, violence, abuse, racial attacks, or any other harmful content. Our goal is to maintain a safe and respectful environment for all users.

How are you looking at local search? Do you rank local web pages differently?

Vinci Mathews: Yes, we do rank local pages in local languages differently. However, one challenge we faced was the scarcity of information in local languages. To overcome this, our Wired Intelligence for Semantic Encoding (WISE) model employs a hybrid architecture where it can read information from English sources and convey answers in local languages.

How will you give clicks to publishers?

Vinci Mathews: We provide a concise summary of information that users can find in each source, without giving detailed answers, and (it) includes proper citations. This approach encourages users to visit the cited sources to read more. Based on our data, we have observed that customers visit at least 2-3 URLs on average for each query. The time users spend on average on a single visit in Nofrills AI is around 8 minutes, while it takes an average of around 20 minutes for users to return after clicking the URL. This suggests that the way Nofrills AI answers — by giving top-priority sources — it effectively encourages users to visit the source URLs.

What is your business model? Are you thinking of tying up with publications?

Vinci Mathews: Currently, our focus is not primarily on a business-to-business (B2B) model, although that option is available. Our model can be integrated into any business framework you have. For example, if you are a publication, you can use our application for various purposes, including internal use. For instance, if you write a story about something that happened 50 years ago and need to look up information, you typically go to Google or your library’s repository. With our API, we can summarise or condense your entire library of information. You can ask specific questions in simple language and our system will provide all the relevant information you need. This can be applied internally within your organisation. Additionally, we can adapt the same technology for customer-facing applications. For example, if someone wants to interact with the publication, whether to place an ad or inquire about products, you can leverage these features.

Google, along with other big tech companies, has faced criticism for its data practices. As data privacy becomes a critical issue in the battle for the future of the internet, what measures have you implemented to ensure that searches on Nofrills remain private?

Vinci Mathews: Privacy is our top priority. We don’t have a history section. This is because we don’t retain queries against users. However, a history feature will be added in the future while maintaining user privacy.