Narrowing the Knowledge Evaluation Gap: Open-Domain Question Answering with Multi-Granularity Answers