Well you can use an LLM similarly. Have it write docs for all your files including a summary for each function / class, ideally in order of dependency. Then use only the summaries in context. This should significantly lower your token count.
Haven't tried it personally but it should work