A controlled experiment compares a top-5 RAG pipeline against a full 127,000-token prompt using the same 12 questions, system prompt, and model (Kimi K3 by Moonshot AI). The goal is to determine whether a one-million-token context window eliminates the need for RAG.

The corpus contains 33 of the author's own articles totaling 127,068 tokens. In the RAG path, the articles are split into 788 chunks of 900 characters with 150-character overlap, embedded with all-MiniLM-L6-v2, and the top 5 chunks (~1,200 tokens) are sent per request. In the long_context path, all 32 articles (~127,346 tokens) are sent every time. Prefix caching keeps the corpus cost at $0.30 per million input tokens versus $3.00 for RAG.

Twelve questions are divided into three difficulty groups: single-fact (Group A), cross-article (Group B), and corpus-wide (Group C). Answers were graded blind—shuffled and labeled X/Y in Excel, with the key file opened only after grading completed.

The experiment reveals that while the long_context approach provides complete information access, RAG remains cheaper and faster for most use cases. Group C questions predictably favor long_context since five chunks cannot cover all 32 articles. However, the key finding is how each model handles uncertainty: whether it admits missing information or confidently guesses.