A knowledge management system has a corpus of 5,000 internal documents totaling approximately 50 million tokens. Users query the corpus for specific information. An architect is deciding between two approaches: (A) stuff all documents into Claude's context window using a very large context model, or (B) implement RAG to retrieve the 5-10 most relevant documents before calling Claude. Which factors most strongly favor approach B?

The retrieval choice that sends only the 5-10 most relevant documents is correct because three factors converge to favor retrieval-augmented generation: first, 50 million tokens vastly exceeds even the largest commercially available context windows, which are commonly in the 200K to 1M token range; second, per-query cost with context stuffing would be catastrophically expensive, because retrieving 5-10 documents reduces input tokens by 99.9%; third, attention dilution is a documented phenomenon where relevant passages buried in massive context receive less attention weight than when they appear in focused context. The always-stuff-everything choice is wrong because even if a context window large enough existed, stuffing irrelevant documents degrades response quality and is economically prohibitive. The structured-versus-unstructured choice is wrong because document structure, whether structured or unstructured, does not determine the appropriate retrieval architecture; token count and cost are the primary factors. The update-frequency-only choice is wrong because while incremental index updates are a benefit of retrieval-augmented generation for frequently updated corpora, the primary reasons to use retrieval-augmented generation, namely token limits, cost, and attention dilution, apply regardless of update frequency.