Question

A developer notices that GitHub Copilot's inline completions in a very large file become slower and sometimes less relevant when they have dozens of unrelated tabs open. A teammate suggests that closing irrelevant tabs and keeping the cursor near related code helps. From an architecture standpoint, what best explains why the amount of context gathered on the client before a request is sent is deliberately bounded rather than including everything available?

Accepted Answer

The underlying LLM has a finite context window, so the client must select and prioritize a limited slice of surrounding code, open tabs, and cursor context to fit the prompt within model token limits.

Answer

The client uploads the entire repository to the proxy on each keystroke, so limiting context only reduces network bandwidth, not model behavior.

Answer

Copilot permanently stores all open-file context as training data, so bounding it exists solely to reduce the size of the training corpus.

Answer

The model retains memory of every prior file the developer edited during the session, so additional context is redundant and discarded at random.