The five stages of RAG

The heart of this chapter is the survey's account of how RAG evolved through five paradigms. Reading them in order shows a clear trajectory: from dumb keyword lookup toward systems that reason about their own retrieval.

Naive RAG

The original, simplest form. It uses keyword-based retrieval methods like TF-IDF and BM25 to fetch documents from a static dataset, then hands them to the model. It follows a plain "retrieve then read" workflow.

Its strengths are simplicity and ease of implementation, which made it a useful proof of concept. But its limitations are real: it lacks contextual awareness (matching words, not meaning, so "car" misses "automobile"), it produces fragmented or generic outputs because it does little preprocessing, and it scales poorly because keyword matching struggles to find the most relevant information in large datasets.

Advanced RAG

This stage adds semantic understanding. Instead of matching keywords, it uses dense retrieval (representing queries and documents as embeddings and matching by meaning), re-ranking (reordering results so the most relevant come first), and iterative or multi-hop retrieval (reasoning across several documents). This makes it suitable for tasks needing precision and nuance, like research synthesis. The costs are higher computation and still-limited scalability on very large datasets or multi-step queries.

Modular RAG

This stage breaks the pipeline into independent, reusable components that can be swapped and reconfigured. Its key innovations are hybrid retrieval (combining sparse keyword methods with dense semantic ones to cover more query types), tool integration (pulling in external APIs, databases, or computation), and composable pipelines (retrievers and generators can be replaced or rearranged independently). The survey gives a finance example: a modular system might fetch live stock prices via an API, analyze historical trends with dense retrieval, and generate investment insights with a tailored model. This flexibility makes it suited to complex, multi-domain tasks.

Graph RAG

This stage adds knowledge graphs to the mix. By storing information as connected entities, Graph RAG can reason over relationships and hierarchies, which strongly helps multi-hop reasoning and reduces hallucination by following real connections. Its limitations are scalability (graphs get expensive to maintain at size), data dependency (it needs high-quality graph data), and integration complexity (combining structured graph data with unstructured retrieval is hard). It suits domains where relationships matter, like healthcare diagnostics and legal research.

Agentic RAG

The final stage, and the subject of the whole survey. It introduces autonomous agents into the retrieval process. Where every earlier stage retrieves in a fixed way, agentic RAG makes retrieval a decision: agents evaluate the query, choose retrieval strategies, retrieve iteratively with feedback loops, and orchestrate the workflow dynamically. Its strengths are adaptability to real-time changes, scalability across multiple domains, and higher accuracy. Its challenges are the familiar costs of agents: coordination complexity, computational overhead from running multiple agents, and scalability strain under high query volumes.