RAG vs agentic AI: a way to ground answers, or a system that acts

AWS describes RAG as a way to make a large language model reference an authoritative knowledge base outside its training data before it generates a response. The point is to answer from your sources rather than from what the model happened to learn.

The approach was set out in a 2020 paper by Lewis and colleagues, Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, which paired a pre-trained generator (parametric memory) with 'a dense vector index of Wikipedia, accessed with a pre-trained neural retriever' (non-parametric memory). The authors named giving provenance for answers and updating a model's knowledge as open problems that this design addresses.

Agentic AI is a different category. IBM defines it as a system that 'can accomplish a specific goal with limited supervision', using agents that can call external tools. RAG settles what the model should know right now; agentic AI settles what the system should do next.

AWS sets out four steps. Documents are converted into embeddings and stored in a vector database. The user's query is embedded and matched against them. The best matches are added to the prompt. The store is refreshed in real time or in periodic batches, so answers stay current.

Every step is fixed. The pipeline retrieves once, whether or not the passages help, and the model writes its answer from whatever came back. That is what makes RAG cheap and easy to test, and also what makes it miss questions that need two lookups.

Anthropic’s Building effective agents treats retrieval as one of three augmentations of a model, alongside tools and memory, and notes that the model can generate its own search queries. In an agentic system, retrieval becomes something the agent chooses to do, as often as it needs, rather than a step that always runs.

Combining the two has its own name, agentic RAG, which has its own page in this series. The point here is narrower: choosing between RAG and agentic AI is usually a false choice. Most agents that answer questions about your organisation use some form of retrieval, and most RAG systems never need to act.

An employee asks how many days of parental leave they are entitled to. A RAG assistant retrieves the leave policy, adds the relevant section to the prompt and answers with a citation. Nothing else happens, and nothing else needs to.

Then the employee asks the assistant to book the leave. That is agentic work: the system checks the employee's record, confirms the dates against the policy, creates a request in the HR system and routes it to the manager for approval. Retrieval still supplies the policy, but the value now comes from the actions around it.

Anthropic’s advice to try a single model call with retrieval and good examples before building anything agentic applies here. If a grounded answer and one click by a person finish the job, an agent adds cost without adding much.

• Use RAG alone when the job is answering questions from your documents with citations, and a person acts on the answer.
• Use an agent when the job ends in an action in another system, such as a booking, a ticket or a record update.
• Use both when the agent’s decisions depend on documents, which is common in policy, support and finance work.

Cortex is a governed AI desktop that answers from your own files. It includes on-device knowledge bases with retrieval, and approvals on tool calls, so it covers the RAG side for a person working with their own documents.

Studio covers the agentic side: agents and workflows that call models through Connect, call tools through integrations and pause at an approval step for a named person. Neither Studio nor Connect replaces your vector store; if your documents already live in one, the agent calls it as a tool.

• Swfte Cortex: Answers from your own files
• Swfte Studio
• AI knowledge base

Facts about other vendors and about the terms on this page were read on the pages below on 7 October 2026. Vendor plans, names and menus change, so check the vendor’s page before you rely on a detail.