Extending Andrej Karpathy's LLM Wiki with 5W1H Framing

This article introduces LLM WikiZZ, an open-source tool that extends Andrej Karpathy's LLM Wiki concept by implementing a structured 5W1H (Who, What, When, Where, Why, How) framing to help Large Language Models (LLMs) better understand the context of a document before answering queries.

Details

The article discusses the 'Transient Knowledge' paradox, where LLMs rediscover documents from scratch for every query, neglecting the 'Context Debt' that builds up when an LLM doesn't truly understand the fundamental frame of the data. LLM WikiZZ is designed to break this cycle by forcing an autonomous 'Discovery Phase' before answering queries, teaching the LLM to architect its own scaffolding. The 5W1H Wiki Frame provides structured context around the target audience, mission, timing, environment, motivation, and formatting requirements, turning raw data into a persistent, shared mental model between the human and the machine. The system includes a 'Contrast Engine' to compare plain queries with the refined 5W1H-framed queries, and an 'Evaluator LLM' to analyze the improvements in terms of situational relevance, concision, and technical depth. The technical architecture prioritizes privacy, security, and persistent context through a zero-server, static-first design and secure CORS proxying.