The fundamental reason is that RAG kind of sucks and requires a ton of effort/optimization to reach a high degree of reliability for most applications. RAG solutions are not Anthropic's core product. Just reading all the relevant files is more expensive but is more effective and efficient from a dev time perspective.