Apple’s CLaRa system fundamentally reimagines RAG architecture by solving the semantic similarity problem that plagues traditional retrieval systems. Standard RAG approaches chunk documents, create static embeddings, and hope queries match semantically similar content—a critical flaw that forces workarounds like HyDE and query decomposition. CLaRa eliminates this by transforming documents into Memory Tokens: compressed representations of pure content stripped of filler and syntax, distinct from raw text. The breakthrough lies in its trainable Query Reasoner. Rather than matching queries directly to documents, it generates hypothetical ideal answers and queries against appropriate Memory Tokens. This component learns from the database itself through training runs, continuously improving retrieval accuracy. The system essentially operates as HyDE amplified—hypothetical document embeddings merged with dynamic, learnable retrieval—while running on modest model sizes. Technical highlights: - Memory Tokens replace static chunk embeddings with compressed content representations - Query Reasoner generates ideal answers before retrieval, not after - Direct database-to-reasoner training loop enables continuous optimization - Achieves superior results without requiring massive computational resources