Agentic Retrieval Augmented Generation (RAG) in Amazon Q Business
Agentic RAG enhances the standard RAG workflow of Amazon Q Business with agentic retrieval and response capabilities. Unlike standard RAG's document retrieval and simple response generation process, Agentic RAG uses multiple intelligent agents and specialized data retrieval tools to deliver more comprehensive and accurate responses while maintaining conversation context.
With Agentic RAG system processes queries through a combination of the following coordinated steps:
-
Analyzes both the user's question and conversation history and determines which retrieval tools to use
-
Intelligently deconstructs complex queries into simpler ones
-
Intelligently triggers multiple data retrieval tools as needed
-
Provides disambiguating questions based on enterprise data to clarify user intent
-
Synthesizes information from various sources, and generates responses with its underlying large language model
-
Provides follow-up questions to intelligently continue the conversation
Throughout this process, it continuously checks response quality and activates additional data retrieval tools when necessary, showing users real-time progress to the user as it processes queries. All responses maintain existing permissions and include clear citations to source material.
Agentic RAG delivers several key improvements over standard RAG. It intelligently selects from available retrieval tools based on query requirements and performs multiple retrieval operations for complex queries. The system maintains conversation context awareness and adapts response generation through retries or disambiguation techniques based on the quality of retrieved information and the subsequent response generated. These capabilities result in higher accuracy, more comprehensive information gathering, and better handling of complex, multi-faceted queries.
To use Agentic RAG, enable the feature using the Advanced search toggle in your web experience interface.
Note
While the Agentic RAG system provides more thorough responses, response times may be longer than standard RAG due to its multiple retrieval operations.