Which llm.generateText feature typically takes longer on the same document, EXTRACT_TEXT or EXTRACT_FIELDS? How does EXTRACT_TABLES compare?