一个 RAG 项目，在真实训练中是怎么被“做出来”的？

Core Insights - RAG technology is not merely a data injection process but a comprehensive framework that reshapes AI understanding and decision-making [1] - The effectiveness of RAG is determined not by the availability of data but by how that data is utilized [2] Group 1: RAG Project Challenges - In conversational AI scenarios, RAG projects face complex tasks that require understanding context, determining useful materials, integrating information, and providing helpful responses [3] - The components of a RAG project include questions, reference materials, and answers, none of which are inherently reliable [4][5] - Common issues with questions include semantic ambiguity, contradictions in context, and illogical leaps [7] - Reference materials may be irrelevant, incomplete, conflicting, or contain common sense errors [8] Group 2: Importance of Human Judgment - The final deliverable of a RAG project is a user-friendly answer, which necessitates meeting specific criteria such as factual accuracy and completeness [9] - Despite advancements in models, significant human intervention is required because 90% of the challenges in RAG projects lie in judgment rather than generation [10][9] - RAG projects train models in three core capabilities: information selection, contextual alignment, and result orientation [11][16] Group 3: RAG as a Long-term Infrastructure - RAG projects are often viewed as transitional solutions, but they serve as a long-term foundational infrastructure in real business applications [12] - RAG acts as a bridge connecting stable models with a changing world, highlighting its ongoing relevance [12][17]