Providing a Large Language Model with 150 tools actually hinders its accuracy, whereas narrowing the selection to just 5 causes relevance to increase sharply. By carefully filtering these utilities, implementing language translation, and adhering to authentic workflows, we successfully lowered our token consumption by 70%.

Read about our approach here:

https://www.civic.com/resources/how-we-built-personal-software-that-adapts-to-you