This version releases the full SheetCopilot proposed and evaluated in our NeurIPS'23 paper.
This version provides the major functionalities of the SheetCopilot agent, including sheet state feedback, error feedback, and external document retrieval.
Apart from the task_sheet_answers which was used as the ground truth solutions in our paper, we also released task_sheet_answers_v2 which contains more functionally correct reference solutions so that we can obtain more accurate evaluation results.
Note that this version still has several technical drawbacks: 1) When extracting the predicted API call from the LLM response, nested quotes in the response may not be parsed correctly; 2) Several APIs' implementations are still not perfect (e.g., FreezePane occasionally fails).
These drawbacks will be fixed in the next version.