I’m building a desktop GenAI code assistant that should:

• Read the user’s entire project folder(potentially thousands of files).

• Provide context-aware code suggestions and debugging help.

• Support multiple LLM providers(OpenAI, Gemini, DeepSeek) using API keys.

Sending the entire codebase in each API request is impractical because:

• LLM context windows are limited.

• Huge latency and network cost if we send megabytes of code each request.

• Cost scales linearly with tokens.

I want to understand how real-world tools like GitHub Copilot and Cursor handle this:

How do they provide accurate code suggestions without sending the whole repository every request?