Builds embeddings from verified samples and knowledge base. Does not modify Ollama base weights — uses prompt-context tuning.