Qianwen Office’s first open source project: distilling the “second you” from Feishu and DingTalk

Zhidongxi Author | Bi Weihao Editor | Li Shuiqing Zhidongzhi reported on August 17 that recently, Qianwen Office has open sourced the personal work context infrastructure MyContext. The project has been online for more than a week and has received more than 1k stars on GitHub. This is also the first open source project of Qianwen Office. MyContext is not a knowledge base in the traditional sense. Instead, it collects chats, documents, meeting records and other work traces in Feishu and DingTalk, and continuously organizes them into a personal work file, allowing the Agent to gradually know who you are, what you are doing, and how you usually deal with things. If you look for a familiar reference, it is a bit like equipping the Agent with a local personal knowledge base Obsidian: it also emphasizes local priority, knowledge organization and data sovereignty. The difference is that the content in Obsidian is mainly written by the users themselves, while MyContext tries to automatically distill it from daily work records. MyContext is actually very much like a very large Loop built around "personal work": messages and work information are continuously captured, then classified, abstracted, stored and associated, and finally precipitated into a context that can be called by the Agent. Looking at the product line of Qianwen Office, MyContext is more like supplementing the "context layer" of Agent Office. Previously, on July 27, Qianwen Office was launched, integrating Alibaba’s three office agent products: QoderWork, Wukong, and MuleRun. On August 3, MyContext was open sourced, providing a practical contextual tool supplement for AI office products. 1. Add context to the office agent and turn the chat record into a "personal file". If you compare the agent to a new employee who has just joined the job, the biggest trouble is not that it can't do the job, but that it doesn't know what you are doing. Every time it takes on a new task, it has to look for clues from prompt words and information again. According to the NANDA report released by MIT, about 95% of enterprise-level generative AI pilots have not achieved benefits. The core reason is the lack of data infrastructure, resulting in A