Andrej Karpathy Revolutionizes Personal Research with LLM-Wiki

Andrej Karpathy, an influential figure in the field of artificial intelligence, recently shared a guide on GitHub to help users build their own personal knowledge base. This guide is part of the LLM-Wiki project, which aims to compile personal notes in a markdown wiki format using an LLM agent.

However, Karpathy emphasizes that existing implementations are often too developer-oriented. To address this, he has developed an alternative approach called DyResearch. This method stands out for its ability to transform a simple coding assistant into a team of specialized agents. These agents, integrated into Obsidian, a popular note-taking tool, combine the concept of a cumulative wiki with lightweight local retrieval through a dual storage architecture.

The DyResearch agents fulfill various roles, each with a specific function:

• Study Coordinator: Manages the organization and planning of research.
• Professor: Provides explanations and teachings on the subjects studied.
• Librarian: Takes care of the organization and classification of information.
• Researcher: Conducts in-depth research on topics of interest.
• Note Taker: Captures and organizes relevant information.

DyResearch is served via a FastAPI backend and connects to Obsidian through a custom community plugin. This system allows for efficient management of sessions and events, as well as optimized retrieval of information sources.

Karpathy also details the technological choices for orchestration and data persistence. For orchestration, he uses Google ADK, while for database and session persistence, he opts for a combination of Postgres with pgvector, as opposed to SQLite local-first and LanceDB. These choices enable users to discuss, ingest documents, and automatically generate or update notes in their vault.