When is a fixed retrieval workflow insufficient?
Conventional RAG often follows one retrieval step with one generation step. Questions spanning documents or systems may need more. Agentic RAG changes the search path as evidence arrives: identify a customer, retrieve a contract and support ticket, then check the latest status. Its purpose is adaptive retrieval, not merely more model calls.
How does an agent choose the next search?
The system interprets the task and chooses accessible sources for an initial search. If an entity or time period is missing, it can seek clarification, rewrite the query, or use another tool. If sources conflict, it checks version, permissions, and authority. Once a stopping condition is met, the model synthesizes the evidence. Production workflows should bound iterations, access, cost, and latency, and retain a record of searches and decisions.
For a question about a customer's delayed delivery, an agent might confirm the promised date in a contract, find the delay in a ticket, and read current status from a project system. Each step reveals what to search next. If the ticket gives no cause, the answer should identify that gap instead of inferring causation from the timeline.
Which questions justify agentic RAG?
Multihop questions, investigations across sources, and tasks that need to adjust search based on intermediate results are good candidates. A direct question answered by one manual may be faster and easier to control with fixed RAG. The agent may decide where to search, while system rules enforce permissions and sensitive actions. MIP's architecture overview provides an enterprise example of coordinating queries, knowledge, and agents.
FAQ
Is agentic RAG automatically more accurate?
No. Additional searches can spread errors unless search decisions and evidence are checked.
Does every task need multiple steps?
No. A simple question should be allowed to end after one successful search.
What should be evaluated?
Assess answer quality, tool choice, evidence coverage, stopping behavior, permissions, cost, and reproducibility.