Retrieval-Augmented Generation

Retrieval-Augmented Generation (RAG) is a technique that grounds large language model outputs in external, retrievable knowledge rather than relying solely on parametric memory. As described in a 2025 Yunqi Capital analysis, its core purpose is to let AI "understand enterprise-specific knowledge and generate valuable output based on it"—addressing the hallucination problem that arises when models lack domain-specific information .

The standard pipeline has three stages: pre-retrieval query formulation, post-retrieval ranking and optimization, and final generation. For structured data like sales records, text-to-SQL paths often outperform vector search; for unstructured documents, embedding-based semantic retrieval is preferred . Research has also pushed into more dynamic forms—Beihang professor Sha Lei's ATM framework, developed with Baidu, uses an adversarial agent pair (an "attacker" that perturbs retrieved content and a "generator" that must produce correct answers regardless) to stress-test and harden RAG systems against ordering and content interference .

At the systems level, Microsoft Research Asia's Baotong Lu has drawn a direct theoretical link between attention mechanisms and retrieval: his "Retrieval Attention" work recasts self-attention as a vector retrieval problem, using CPU-GPU heterogeneous architectures to offload KV cache storage and dynamically fetch only relevant context—cutting memory use by roughly 90% while maintaining near-full precision on most tasks .