I am working on a game https://imperiaquiz.com/en that needs over 20,000 questions. I use jev to classify 'has statement' and 'can be a question' paragraphs/chunks taken from cli script to reduce token usage. Meaning, an LLM only starts work once I've chunked text and marked it as 'to be reviewed' instead of parsing the full content. This reduces tokens usage at least 10x on average