Compare long context, RAG and a router on your questions

The three strategies share one set of answer rules and one schema, and the context always goes in a system-prompt file in front of the question. Only the context changes: the whole bundle, 20 chunks, or 20 chunks then whole documents. The decision that matters most is putting the bundle in --system-prompt-file with byte-identical contents on every call. That makes the bundle a stable prefix that can be cached, keeps it off the command line, and lets you confirm caching from cache_read_input_tokens instead of assuming it. The RAG retriever's dense half goes through the OpenAI SDK (OpenRouter or Ollama), or a fake embedder when .env is absent. Every answer runs on your subscription.