RAG agentique
Classic RAG always retrieves: before every answer the pipeline fetches documents, whatever the question. In agentic RAG that retrieval becomes a decision the agent makes. It judges whether it needs to search at all, picks the source, reads what came back, weighs whether its draft answer holds up, and runs a different search when the first one returned nothing useful. A fixed pipeline step becomes a loop with a policy. What it costs: a call count and a latency you can no longer predict before execution.
Strengths
- The agent can answer without searching when the question is trivial, which a fixed pipeline cannot do
- An insufficient answer can trigger a second, differently phrased search instead of shipping as is
- Arbitrates between several heterogeneous sources, something a single retrieval step cannot express
Limitations
- Cost and latency turn unpredictable: you no longer know before execution how many calls a question will trigger
- Amplifies upstream defects instead of fixing them: bad chunking gets badly retrieved several times over
- Far harder to debug: the failure sits in a chain of decisions, not in a query you can replay
Best for
- Several bases of different natures where each question genuinely requires choosing between them
- Open questions where a single query never gathers the whole answer