RAG Isn't Dead — Your Chunking Is

The obituary was premature

Every long-context release re-runs the "RAG is dead" cycle. Then the bill arrives. Retrieval remains how you make LLMs cheap, current and grounded.

Five strategies, one winner (it depends)

• Fixed-size — fine for homogeneous text, terrible for code.
• Recursive — the sane default.
• Semantic — best quality, highest cost.
• Document-structure — unbeatable for markdown/HTML.
• Agentic chunking — an LLM decides boundaries; expensive, occasionally magical.

Benchmark on *your* corpus. The winner changes with document shape more than with embedding model.