Agentic RAG adds planning and iterative retrieval to retrieval-augmented generation. An agent can break a question into parts, inspect available results and search again to gather relevant evidence for an answer.
Use more than one retrieval step when needed
Some business questions span several documents or depend on conditions that are unclear at first. A simple search-and-generate workflow may miss this context. Agentic RAG can adapt the retrieval process as information emerges, rather than relying on a single fixed search.
The approach
The agent chooses among configured retrieval methods and tools, evaluates the available information and refines the query where useful. Research on agentic RAG describes planning, tool use and reflection as possible design elements; the selected approach depends on the task.
- Plan: Identify what the question requires and which sources may help.
- Retrieve: Use permitted keyword, vector or graph search and configured tools.
- Assess and refine: Look for gaps or conflicts and repeat retrieval within defined limits.
- Respond: Produce an answer with available source references, or indicate that information is insufficient.
Business value
- Gather broader context
Combine evidence from multiple sources and conditions.
- Support source checks
Keep references so users can examine the material behind an answer.
- Reduce manual search steps
Automate parts of moving between permitted sources and refining queries.
Application considerations
- Repeated retrieval can add latency and cost. Configure iteration limits and simpler paths for questions that do not need multiple steps.
- Agent self-evaluation and source citations do not guarantee correctness. Test handling of missing evidence and contradictions, and restrict tools and data to the permitted scope.
Related products
- AIKNOW
Uses agentic retrieval to connect enterprise knowledge with decisions and authorized actions.
- AIKNOW Answer
Combines conversation context and business procedures for configured customer-service automation.
References
- Lewis, P. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020. arXiv:2005.11401. https://arxiv.org/abs/2005.11401
- Singh, A. et al. (2025). Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG. arXiv:2501.09136. https://arxiv.org/abs/2501.09136
- Yao, S. et al. (2023). ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023. arXiv:2210.03629. https://arxiv.org/abs/2210.03629