Karpathy's LLM Wiki concept just got a real implementation. @lumimaxAI built a deep knowledge engine that actually works.
Core architecture differences:
1. True knowledge graph persistence. Doesn't re-parse raw files on every query. Builds a structured KB that LLMs query directly. Think vector DB meets semantic graph.
2. Citation tracing at the source level. Every response links back to exact pages in original docs. Verifiable retrieval, not hallucinated summaries.
3. Conflict detection layer. Surfaces where sources contradict each other instead of averaging them out. Critical for research workflows where ground truth matters.
4. MCP protocol support. Your KB becomes a universal context provider. Any MCP-compatible agent (Claude, custom tools, etc.) can hook into it. Build once, reuse everywhere.
Tradeoff: Heavy upfront processing time for large document sets. But that's the cost of actual semantic parsing vs. shallow embedding.
Still requires human judgment on conflicting sources. The system flags disagreements but doesn't resolve them. Feature, not bug.
Core architecture differences:
1. True knowledge graph persistence. Doesn't re-parse raw files on every query. Builds a structured KB that LLMs query directly. Think vector DB meets semantic graph.
2. Citation tracing at the source level. Every response links back to exact pages in original docs. Verifiable retrieval, not hallucinated summaries.
3. Conflict detection layer. Surfaces where sources contradict each other instead of averaging them out. Critical for research workflows where ground truth matters.
4. MCP protocol support. Your KB becomes a universal context provider. Any MCP-compatible agent (Claude, custom tools, etc.) can hook into it. Build once, reuse everywhere.
Tradeoff: Heavy upfront processing time for large document sets. But that's the cost of actual semantic parsing vs. shallow embedding.
Still requires human judgment on conflicting sources. The system flags disagreements but doesn't resolve them. Feature, not bug.