LLMs have context limits. Embeddings work best on focused text. You can't embed an entire 50-page document and expect the vector to capture every idea — the meaning gets averaged into mush. Chunking breaks documents into retrieval-sized pieces that each carry a clear semantic signal.

The chunking strategy you choose directly impacts retrieval quality. Get it wrong and your RAG pipeline will retrieve irrelevant content no matter how good your embedding model is.