They are significantly better! IDEs aren’t just like LLMs; they transform your codebase into embeddings (knowledge), giving them a much greater advantage in understanding larger contexts and effectively acting on the code. Plus, they come with proper tool integrations, allowing them to apply changes seamlessly while including fallback mechanisms to handle any issues during the process.
>they transform your codebase into embeddings (knowledge), giving them a much greater advantage in understanding larger contexts and effectively acting on the code.
Is this true? Isn't the context size controlled by the model? Is there any difference (outside convenience) to pasting your code in the web interface?
Sure thing! They can’t just dump all the code from your repository into the context.
The first step is identifying the relevant code in your repo, likely using an indexed vector database. They probably transform your question or prompt into an embedding as well to enhance accuracy. Once they’ve pinpointed the relevant sections, they send those specific parts of your code to the LLMs for processing. I actually wrote an article diving deeper into this concept—check it out!
https://medium.com/@level09/cursors-secret-sauce-the-embed-t...