Hello, Hello! Since 2023, I’ve been testing new AI models like OpenAI's GPTs, aiming to create new financially viable and functional solutions.
For example, GPT 3.5 already delivered good results, but with high costs, so we continued developing solutions like our Workflows and Caramelo AI Marketing to generate content with maximum token efficiency.
About two months ago, the mini and flash versions of the models were released—models with much lower costs while maintaining good response quality. With these models, creating AI-driven solutions became much more feasible.
After their release, we started testing and developing Fina, initially using part of the logic with buttons, elements, and conventional chatbot flows. The big problem here was that, to be a true assistant, we would have to add many features specifically programmed for each use case—something unfeasible given the available investment.
Users were asking unexpected questions, "breaking" the application, which didn’t react well. Back then, things like "Petrobras' financial health" didn’t work, but they were interesting to the target audience.
Since I already knew about AI agents, I saw the perfect opportunity to apply that knowledge to Fina and make her much smarter.
The team chose LangChain as the framework for this version, and after testing a few libraries, we went with LangChain in Javascript.
We moved forward with development using LangChain in a RAG format (Retrieval-Augmented Generation), where the AI decides what data it needs and how much of it. We have expertise in financial data and accessing official data sources like CVM, which helped improve the solution’s final quality.
The biggest challenge was ensuring that the AI understood that certain data belonged exclusively to a specific company. With some prompt engineering techniques, we solved this issue.
Today, Fina provides accurate and reliable data with great insights and has a lot of autonomy in deciding what to do when receiving user messages.
About AIs and the current state:
AI still makes mistakes and needs assistance to handle complex situations, but with some adjustments, their reasoning power (yes, they have a certain level of reasoning, and it’s impressive) can meet an incredible number of user cases—simply by giving them the right tools to act while ensuring the necessary limitations and precautions.
I've been developing automation solutions for years—many of which used to take hours to build. Today, AIs simply do tasks like formatting variations of a "humanized" text. It’s inevitable that we’ll have AIs for practically everything, and society needs to prepare for that…
Whether generative AI models think or not doesn’t really matter today. What matters is testing and understanding what this tool can do, and using it to create new possibilities.