I recently came across a weekend project shared by Michael Vicente, and it changed how I think about building AI-powered applications.
Michael built a Retrieval-Augmented Generation system for AIO Growth, connecting ChatGPT to a database containing more than 5,000 AI tools. The goal was to provide fast and personalized tool recommendations without sending the entire database to the AI every time someone submitted a question.
At first, the idea of connecting AI to thousands of tools sounded like it would require sending a massive amount of information with every request. That would consume more tokens, increase costs, and potentially make the application slower.
Michael approached the problem differently.
Instead of giving ChatGPT access to everything at once, his system follows three focused steps.
First, ChatGPT identifies what the user is looking for. Next, MongoDB searches the database and retrieves only 15 to 20 relevant tools. Finally, ChatGPT uses that smaller selection to create personalized recommendations.
User submits a question
↓
ChatGPT understands the intent
↓
MongoDB retrieves relevant tools
↓
ChatGPT creates personalized recommendations
What caught my attention was not only how the system worked, but also the results Michael reported.
His approach reduced the cost per query by 93 percent, from approximately $0.0008 to $0.00005. It also made responses 40 percent faster, with an average response time of around 1.2 seconds.
He used GPT-4o-mini for cost-effective reasoning, MongoDB for semantic filtering, and compact tool summaries that reduced token usage by approximately 60 percent. Guests receive access to Smart Search, while signed-in users can use the full conversational AI experience.
The biggest lesson I learned from Michael’s discovery is that building a smarter AI application does not always mean sending more information to the model.
It means sending the right information.
Before learning more about RAG, I thought an AI system would provide better answers if it received as much data as possible. Michael’s project showed me that too much unnecessary information can increase costs and slow down the experience.
Retrieval makes the process more focused. Instead of asking the AI to search through thousands of possibilities, the system first selects a small group of relevant options. The AI can then concentrate on explaining and personalizing those results.
I also learned that building with AI is not only about choosing the most powerful model. The structure around the model matters just as much.
The database, retrieval process, summaries, access levels, and user experience all contribute to the final result. A smaller and more affordable model can still deliver a strong experience when it receives relevant and well-organized context.
Michael’s project also reminded me that useful technology often begins with a practical question:
How can this be made faster, more relevant, and less expensive?
As someone who is currently learning coding and automation, this gave me a clearer understanding of how AI applications can be designed for real-world use.
It is not simply about connecting an application to ChatGPT. It is about deciding what information the model needs, when it needs it, and how to deliver that information efficiently.
Michael’s project made RAG easier for me to understand. More importantly, it showed me that optimization is not always about adding more powerful technology.
Sometimes, the smartest solution is reducing the amount of work the AI needs to do.
Top comments (0)