Personal Wiki Agent: A Local Multimodal Knowledge Assistant Based on Deep Agents

Personal Wiki Agent is a personal knowledge assistant for Feishu/WeChat scenarios, built on the Deep Agents framework. It integrates lightweight intent recognition, multimodal RAG retrieval, and dedicated sub-agents to achieve a complete closed loop from knowledge precipitation to intelligent Q&A. It supports local deployment to protect privacy, unifies indexing of knowledge from multiple sources such as Feishu Wiki, documents, images, links, and Xiaohongshu posts, and provides personalized intelligent Q&A services.

In the era of information explosion, personal and team knowledge assets are scattered across Feishu documents, Wikis, web links, screenshots, Xiaohongshu notes, etc., making it difficult to retrieve and utilize them quickly. Traditional search only uses keyword matching and cannot understand semantics; cloud-based large model solutions have data privacy and cost issues. The project aims to complete knowledge indexing and retrieval locally, balancing privacy protection and intelligent Q&A experience.

Multimodal Knowledge Indexing

Uses Qdrant vector database to locally index multiple sources such as Feishu Wiki/documents, web links, image content, and Xiaohongshu posts. All content is vectorized locally to ensure data privacy.

Lightweight Intent Routing

Built-in intent recognition model to distinguish between knowledge retrieval, precipitation, chat, etc., for efficient use of system resources.

Dedicated Retrieval Sub-agent

Handles retrieval requests: converts questions into query vectors, performs Qdrant similarity search, reorders and filters results, and organizes context for the generation model.

Generation and Q&A

Sends the retrieved context and user's question to the large model to generate personalized and relevant answers.

• Local Deployment: Can run on local servers/devices; sensitive data does not leave the local environment, protecting privacy.
• Multimodal Understanding: Supports image content extraction (e.g., text and charts from meeting whiteboards) and incorporates them into the knowledge base.
• Feishu Ecosystem Integration: Natively supports Feishu group/one-on-one chat bots, allowing direct interaction in chat windows.
• Progressive Knowledge Accumulation: Precipitates knowledge by forwarding links, pasting text, uploading images, sharing Xiaohongshu posts, etc. The more the knowledge base is used, the smarter it becomes.

• Personal Knowledge Management: Unifies indexing of bookmarked web pages, notes, and screenshots; obtains answers via natural language questions.
• Team Knowledge Base Q&A: New employees can understand company policies and project backgrounds through conversational interactions without flipping through documents.
• Meeting Content Precipitation: Stores meeting whiteboard photos and audio-to-text transcripts; allows querying discussion content at any time.
• Cross-platform Information Integration: Connects Feishu, WeChat, Xiaohongshu, and other platforms to establish a unified knowledge entry point.

The project is developed based on the Deep Agents framework, which provides basic capabilities such as agent orchestration, tool calling, and memory management. The RAG pipeline uses embedding models to convert text and images into vectors, which are stored in the Qdrant vector database. During retrieval, approximate nearest neighbor search is used to find relevant content, and then a reordering model optimizes the result quality.

Personal Wiki Agent demonstrates a solution combining large language models with local knowledge bases, providing intelligent services while protecting privacy. Its design of multimodal indexing, intent routing, and sub-agent collaboration provides a practical technical solution for building personal/enterprise second brains. As multimodal model capabilities improve, local knowledge assistants will play a role in more scenarios.