From what I can see, they download the repo to their server, generate per-function semantic embeddings (which will no doubt cost them at scale) [1], then use some combination of GPT-3 and Codex prompted with something like "Here's the source code of a git repo. Answer the following questions about it: {USER_QUESTION}."
[1] https://beta.openai.com/docs/guides/embeddings