kannthu

A few months ago, Cursor released a technical blog post describing how they implemented the instant apply feature. The post went deep into how and why it works so well. They even did a part 2 in collaboration with an LLM inference provider - https://fireworks.ai/blog/cursor.

However, when I checked today, the blog post had been deleted. If you go to the https://www.cursor.com/blog/instant-apply, you will get 404.

Are they scared of the competition? The PearAI started getting some traction recently.

BTW the blog post is still accessible on the Wayback[1] + it seems like somebody already implemented the instant apply and published it on Github[2].

[1] https://web.archive.org/web/20240823050616/https://www.cursor.com/blog/instant-apply

[2] https://github.com/llllvvuu/instant_apply

*TL;DR;*

Over one month ago I posted about a really hard problem that I "accidentally" solved (https://news.ycombinator.com/item?id=40460084).

The problem is to resolve cross-file references for multiple programming languages. I can generate a graph representation of the codebase.

*Why do you need to have a graph representation of the codebase?*

- To understand how code references other code

- Track how data is passed around

I generated references for repo https://github.com/dj-stripe/dj-stripe, here is a gist: https://gist.githubusercontent.com/kannthu/6e1bdd2781d2e0a6ded30844d61f089e/raw/f1fa4bc0f34891834ce13ac256eec12f6cc671e1/dj-stripe-references.json

The gist is a big JSON blob that contains definitions form the repository.

Definitions are:

- top-level functions

- classes

- methods and public properties

- top-level variables

- exports

Each definition contains:

- Snippet, path, and range within the file

- "references" - a list of places where the definition is used

- "expressions" - a list of resolved references (variables, functions, and classes) that are used within the body of the definition

*How this data can be useful?*

If you are building code generation, code intelligence, or code review products - your product needs to have an understanding of the codebase for many programming languages at once. The more accurate context you feed to LLM => the better output you will get, and doing it in-house is really expensive and resource-consuming.

Let me know if it is interesting for any of you.