This project can be found on GitHub
Since the craze of LLMs started I have kept seeing different projects and posts for “talking to XYZ”. Whether it be talking to a database, a PDF, or a codebase. That last one of talking to a codebase always interested me, and now with context windows for models being so large the contents of a moderate sized codebase could probably be stuffed into one.
This was really appealing to me as I often have to work with new repos or analyze open source code and try to figure out what is happening or what they are doing to get something done. Then I realized that I’m not actually sure how people are talking to their codebases. I’ve seen a couple of tutorials that talk about doing a RAG query where they embed the files in a codebase and then let a model query it to add to the prompt. However, I want to simply add everything to context.
So I wrote a little script that will format the contents of a git repo into text that I can copy and paste into an LLM chat window.
At the time of developing this script Anthropics Claude Sonnet 3.5 was the best model I found and had a context window 200k tokens. In the prompting docs they also recommended using xml tags to structure prompts. So I figured I can use the xml tags to communicate the directory structure of a repo.
The approach is really simple of taking every file in a locally cloned repo and appending its contents to a string. Each file’s contents would be in between xml tags that communicated the file path e.g.
<README.md>
Lorem Ipsum Dolor
</README.md>
<lua/plugin/init.lua>
function init()
end
</lua/plugin/init.lua>Directories would be shown by creating parent xml tags.
After the final prompt is constructed I also use a third-party library tokencost to quickly check the number of tokens produced and the cost of the prompt.
The script also has a set of hardcoded directories and files that it ignores
when making the prompt such as a LICENSE file. These are viewable in the
main.py script in the repo
Further Development
This was a really quick and dirty script I got working for analyzing a few repos one day. There’s plenty of room for improvement
- Supporting relative file paths
- Supporting remote repositorys and not requiring a local clone
- Respect
.gitignorelists for ignored paths