Inside the First Advanced Research & Reasoning System for LLMs: How You.com Built an AI That Thinks Like a Research Analyst

This summer, we embarked on the challenge: How could we make our Research Agent 10x better? Our Research Agent was already helping millions of You.com users with research by searching for and synthesizing information into well-cited, comprehensive reports. But we wanted to go further.

Today, the result is the first Advanced Research & Reasoning system that transforms LLMs from simple question-answerers into sophisticated research partners. Enabling AI to analyze up to 200 sources per query and employ PhD-level research methodologies represents a shift in how AI conducts research.

Compressing learning: from undergrad to PhD with practical experience and understanding

Traditional large language models are like highly intelligent and motivated undergrads—full of potential yet lacking the depth of practical expertise or the understanding of full contextual correctness. And, when they don’t know the answer, they’ll very confidently express that they do. When we connected LLMs to real-time news and web results, we were able to empower them to validate their own answers. 

But what comes next? As a recent MIT/Harvard study showed, LLMs by themselves lack a coherent understanding of the world. If you think about the difference between an undergrad and a PhD, a PhD has specialized in a topic by reading all the relevant research, deeply considered the hard questions in the field, and has the skills to explore the unknown areas of their field. 

With our latest feature, Advanced Research & Reasoning, we’ve taken LLMs to a whole new level—think of it as enrolling LLMs in a PhD program; it’s now equipped with years of experience, expertise on a topic, and an instinct for what the important and insightful questions are. Our new Advanced Research & Reasoning feature does the same and enables AI to function as an intelligent collaborator rather than just an answer generator.

Mirroring human learning: the blueprint

We architected this solution around the way highly skilled and intelligent people break down complex problems into multiple steps where learnings from each step inform the next. 

Imagine you’re a consultant or research analyst tackling a new strategy problem - what would you do? You’d start by grounding yourself in data (aka reviewing the current leading theories, news, and findings), then formulate thoughtful questions or hypotheses and do more research on the deeper topics. 

We envisioned the AI as a diligent student enrolled in a PhD-level class, doing its homework before crafting a final thesis. This rigorous approach ensures that no matter what users ask, they receive responses that are not only accurate but also deeply informed.

The benefits of this reasoning capability: your AI PhD co-worker

So, what can you expect from this game-changing feature? Instead of simple back-and-forth exchanges, you can present a “challenge” to the AI, which will:

• Break down requests: The AI dissects queries into component parts, understanding the essential data points needed for a comprehensive response.
• Develop a research plan: Once the request is clearly defined, the AI crafts a plan and undertakes extensive research using real-time information, drawing from live web sources or specified files.

This systematic approach not only enhances the quality of responses but also facilitates a meta-synthesis of the gathered information, culminating in nuanced and insightful replies.

Why this matters: the future of research is here

The landscape of AI is crowded with tools that merely generate answers without a solid grounding in data. Some rely on basic web searches, while others dabble in rudimentary internal reasoning. You.com’s Advanced Research & Reasoning mode combines these capabilities, offering a robust solution that performs deep research across selected sources—potentially utilizing up to 200 sources per query. 

You can rely on the Advanced Research & Reasoning feature to:

• Build trust: By showcasing its work through mini-reports, the AI allows users to verify and expand their understanding of complex topics.
• Be customizable: The feature can be adapted and enabled on any Custom Agent, providing flexibility in research workflows.

Here are a couple example applications of this capability in real-world scenarios

1. 1. A comprehensive analysis of the local commercial real estate market
   2. Financial overview for a major institution like Scotiabank

These use cases highlight how Advanced Research & Reasoning can transform data into actionable insights. You can try out the advanced reasoning capability now in You.com’s Research Agent, and later this month in all of our agents, including Genius and custom agents. 

Lessons learned: Insights for product development in the world of GenAI

We’ve learned invaluable lessons along the way, particularly in working with LLM's in the development of new products and we think they are relevant for your product development projects as well:

• Build conviction through iterative component development: Each stage of the process allowed us to build conviction in our plans by prototyping components along the way. First, we built Custom Agents. Once we had those, we used Custom Agents to prototype the next components in the workflow itself. Stepwise incremental development milestones work the best.
• Use GenAI as your development partner: As we trained and fine-tuned our systems, we used GenAI as to judge and evaluate responses. It would’ve been impossible to deliver this capability in the time and with the resources we had; we needed GenAI to handle what would have taken human experts literally thousands of hours.
• Anchor in real-world use cases: Staying close to user feedback is a product truism. But even more so in the field of GenAI, where the potential to do anything (and the temptation to build anything for everyone) is real. The emphasis on real-world use cases helped anchor our development. It’s easy to prototype; it’s hard to do a specific, real-world task well.

The future means partnering with AI

The opportunities for product innovation aren't simply better or faster models, but building enhanced capabilities on top of existing and new models when they appear. And, building with GenAI at every step of the process, from ideation to thought partner to evaluator.

At You.com, we believe in harnessing the power of GenAI to unlock human potential. Our Advanced Research & Reasoning feature empowers you to build your own Custom Agents, breaking down complex analyses into manageable parts, and saving time for the relational, interpersonal, and creative work that only humans can do. We’re excited to redefine the future of human-computer research and reasoning—one insightful query at a time. 

Reach out if you’d like to chat more about product development in the world of GenAI, especially if you’re looking for production-ready APIs for LLMs grounded in real-time news and web, or our latest enhanced GenAI features, like Advanced Research & Reasoning.