Well, theoretically you can give it up to the context size minus 4k tokens, because the maximum it can output is 4k. In practice, though, its ability to effectively recall information in the prompt drops off. Some people have studied this a bit - here's one such person: https://gritdaily.com/impact-prompt-length-llm-performance/