Code Execution with Gemini | Intro to Tools

Key Takeaways

Large language models have come a long way when it comes to solving math problems, but fundamentally they still work by predicting the most probable next word in a sequence based on patterns learned during training. In other words, they aren't calcs and they aren't the most efficient or accurate at numerical computation. But by connecting large language models to tools, we can extend their abilities. So, let's check out the code execution tool, which makes Gemini more effective at problems that involve math, data analysis, computation, and code-based reasoning. Okay, I'm going to jump into Google AI Studio, and I'll make sure I have Gemini Flash latest selected as my model, and we'll type calculate the sum of the first 60 prime numbers. And I did check this beforehand. The answer is 7,699, but let's see what Gemini says here. Oh boy. All right, we got 8,368. So, I've run this query a bunch of times. Um, sometimes it works, sometimes it doesn't. Um, the model has a whole lot of hard work that it's put into here to try and solve the problem. But again, to come to the answer, it's really just using this next token mechanism um predicting the most likely next token. And that's not the best way to solve a math problem like this. So, we can use the code execution tool to give Gemini a little bit of help here and improve the ability to come up with a correct answer. So, what I'm going to do is I'm going to copy this prompt and let's restart. I'll type it in again here, but this time I'll make sure I have code execution selected in the tool toggle over here and we'll run the prompt. And this time, you can see we actually did get the correct answer, 7,699. But instead of that kind of like long string of of text that the model had produced earlier trying to come up with an answer, it's returned Python code. So while LLMs maybe aren't the best at doing actual mathematical computation, we know they're really really good at language and they're really good at generating code. So we can generate the code to solve the problem, which in this case is a function that checks if a number is a prime number or not and then adds up um the first 60 prime numbers. And this Python code is executed um on the API back end and it's returned and the result is returned to you right here which is 7,699. So not only does this make the model better at doing math, but it also gives us this nice chunk of code here where we can always refer to and see, okay, where did this answer actually come from? Okay, so let's check out a more practical example of how you might use code execution um in the real world. So, here's a little app that I made with uh Vibe Coding in AI Studio and uh it's my CSV Genius app. So, we're going to upload a file here and we'll say what fruit did I sell the most of parenthesis quantity. Okay, so let's see what this is. So, while the model's thinking, here is the uh CSV file that I uploaded. It is data from my tropical fruit stand that I wish I indeed owned. Um so there's a transaction ID, the date, the product sold, um quantity, price, etc. So the idea here is can we ask questions about this data um things that you would sort of normally as a data scientist maybe try and solve in pandas. Um but this makes it a little bit easier if you don't know how to use pandas and you don't know how to um do this kind of data analysis. We'll let the model do it for us. So, we'll jump back into the app here, and you can see it says that I sold the most of mangoes, and that was 63 units. So, that's all fine and great. Uh, but now I'm wondering, well, how did it actually come up with that answer? Um, so I have this little generated code toggle right here that I've added in. And we can see the model's loaded in the data. And then once it's done that, here is the little code snippet that Gemini wrote that was then executed. So, it loaded in my data as a data frame. It filtered for fresh fruit and then um eventually prints out the fruit that I sold the most of. So now we know where this answer came from. It was this code that was generated by Gemini that was executed on the back end and that answer was returned to um returned to the user and that's where Mango came from with 63 units. If we toggle over to the code section, you can see here in the code uh during the AI chats create method um we have Gemini 2.5 Flash selected as our model. And then we also have the code execution tool uh being passed here. And so that is how the model is able to call out to that tool to run the Python code that it generates. And that is all for code execution. But if you build something cool with this tool, definitely let me know.