A New AI Search Engine Is Challenging Perplexity. And It’s Decentralized.
AI-driven search is difficult to fathom from just a few years ago, and yet it has already become a daily part of our lives. We use it for casual search and in depth research alike, trusting AI to find what we need, provide us with well constructed answers, and do all of that with our sometimes overly complex questions. With so many newcomers in this industry, how can we tell when search algorithms are the best? This is a question that Google aimed to solve, and their solution was the FRAMES evaluation dataset, designed specifically to test Retrieval-Augmented Generation (RAG) systems.
This setup can actually provide insight into just how impressive these search engines can be, as the FRAMES set requires a search engine to first understand a user’s question, be able to dive into the internet and search for the different elements of the search query, then organize each of the pieces into an answer that is both coherent and correct. While this is what humans do when we research an answer, the amount of understanding required for subtle differences in meaning, context, and just poor grammar constitutes a mountain of knowledge needed to accomplish this. How well do AI-driven search engines perform? Honestly, a lot better than you’d think.
Also Read: AiThority Interview with Brian Stafford, President and Chief Executive Officer at Diligent
According to an early story on the FRAMES announcement, the data set includes over 800 “multi-hop” questions, requiring the engine to “hop” across different sources and retrieve information from 2-15 different documents to get the full answer, then synthesize this information into a proper answer. When only allowed a single step, the best engines could achieve a 40% accuracy, while allowing them to access all the data needed jumped up the accuracy to over 70%. This says a lot about helping engines get the data sources they need to answer correctly, but also shows that even when limited, these engines show significant promise for such difficult queries. It stands to reason that the billions put into these AI-driven search engines are required to compete at such an elite level, being able to train with massive documents and having advanced algorithms to process similar to a human.
Given this, imagine the shock when these score records were soundly broken…by an open source, decentralized search engine.
Sentient Chat: Open Source, Decentralized Knowledge
The AI firm Sentient, focused on helping its clients build their own AI Models, recently took on the FRAMES dataset and tested the results of its new platform, Sentient Chat. The results of the test were published, but a few elements of this platform are different from competitors. Sentient Chat is decentralized, meaning that it is not driven and maintained by a single team, but rather by contributions from countless participants across the globe. Each provides a level of computing power and validation so that the search engine can live across the network of participants, not within a single (and vulnerable) server. This helps keep the engine censor-proof, but also helps a great deal with scalability, as a larger following means stronger computation and capability, without the need for expensive, step-wise capital expenditures.
Sentient Chat is an agentic chatbot, and has over 15 agents natively integrated. Based on its early performance, the app had over 1M users on its early access waitlist before it went live on February 26. The core of the chatbot is Dobby, the world’s first community-owned AI model. “Community-owned” here is exactly what it sounds like: over 660k users were given access earlier this year, and holders of Dobby NFTs (tied to Sentient’s flagship Loyal AI model) were given priority access to the release of Sentient Chat.
While decentralized AI engines are novel, Sentient Chat isn’t the only one out there; notably, Perplexity AI has performed well in many aspects of decentralized search. However, Sentient Chat has one more trick up its sleeve: it is not only decentralized, it is open source.
🌐𝐀𝐍𝐍𝐎𝐔𝐍𝐂𝐈𝐍𝐆 𝐒𝐄𝐍𝐓𝐈𝐄𝐍𝐓 𝐂𝐇𝐀𝐓—𝐓𝐇𝐄 𝐎𝐏𝐄𝐍, 𝐀𝐆𝐄𝐍𝐓𝐈𝐂 𝐏𝐄𝐑𝐏𝐋𝐄𝐗𝐈𝐓𝐘
𝐎𝐩𝐞𝐧-𝐬𝐨𝐮𝐫𝐜𝐞 𝐀𝐈 𝐢𝐬 𝐡𝐞𝐫𝐞.
𝐉𝐨𝐢𝐧 𝐭𝐡𝐞 𝐰𝐚𝐢𝐭𝐥𝐢𝐬𝐭 𝐟𝐨𝐫 𝐢𝐧𝐯𝐢𝐭𝐞 𝐜𝐨𝐝𝐞𝐬:
👉 https://t.co/VcTTWEhPz3Perplexity has done great work in the… pic.twitter.com/ysaXkLOPU3
— Sentient (@SentientAGI) February 26, 2025
The ability for anyone to see how the logic of its search works, and the ability to create your own model has evolved into the ability for niche AI agents to be developed, trained, and deployed quickly and easily. Researchers can more easily examine what is happening “under the hood,” allowing for more rapid and effective results.
Speaking of results, how well did Sentient Chat perform using the FRAMES evaluation? Based on the 2024 results, the search engine did exceptionally well. For the overall benchmark, the best performance of an open source system up to that point was 34.1% accuracy, with a closed system performance by GPT-4o achieving 50.5% accuracy. Sentient Chat however, successfully achieved a massive 56.7% score. When testing against the SimpleQA accuracy using multiple search queries, Sentient Chat outperformed all other models again, even squeaking out a victory over the distributed, but closed source, Perplexity model with a 94% accuracy over Perplexity’s 93.9%.
Also Read: How the Art and Science of Data Resiliency Protects Businesses Against AI Threats
What Does It All Mean?
These numbers seem pretty great, but what do they actually mean? For the average user, having a high score is of course nice because it represents a better model and more likely, better results. However, there are bigger implications with a model like Sentient Chat performing so well. This is good for the average user for several key reasons. First, the decentralized nature of the search engine means that as it gains popularity, its capabilities and capacity to handle many different queries will only grow. As more participants want to take part and earn rewards for their participation, the engine instantly gains more computing power and can run even more cost effectively. Second, the open source nature of the search engine is already attracting major attention from developers, and a waiting list is filling up extremely quickly as developers see countless use cases for a highly accurate search engine that can operate with dense, complex search queries with consistent results. The benefit for the average person is that as more and more developers are using this model, the many different apps they create will come complete with a high performance search capability that is structured to continuously improve through infrastructure expansion, and through various innovations from developers making their way back to the main platform as a best practice.
It’s certain that the other players in the search engine market will continue to improve their own models, and improve the scores from FRAMES and other dataset tests. However, the achievement from Sentient Chat shows that rather than being a disadvantage, a search engine that is both decentralized and open source has a great deal to offer, and will be able to keep up with closed systems funded exponentially more.
Comments are closed.