Why traditional RAG was not enough

The survey spends time on the specific limitations that pushed the field toward agentic RAG. These are worth understanding because each one is a problem an agent is meant to solve.

Contextual integration. Even when traditional RAG retrieves the right documents, it often fails to weave them into a coherent answer. The survey's example: a query about the latest Alzheimer's research and its implications for early-stage treatment might retrieve relevant papers, but a static system fails to synthesize them into an explanation connecting new treatments to specific patient scenarios.

Multi-step reasoning. Many real questions need multi-hop reasoning, retrieving and synthesizing across several steps where each depends on the last. Traditional RAG cannot refine its retrieval based on intermediate insights. The example: a question about applying European renewable energy policy lessons to developing nations, including economic impacts, requires orchestrating policy data, regional context, and economic analysis, which a static pipeline cannot stitch together.

Scalability and latency. As the volume of external data grows, querying and ranking become computationally heavy, causing latency. In time-sensitive settings like financial analytics or live support, that delay undermines the system's usefulness.