Amazon announced agentic retrieval for its Amazon Bedrock Knowledge Bases, enabling models to break complex questions into sub-queries for more precise answers. This feature is available on Amazon Bedrock Managed Knowledge Bases, which handles chunking, embedding, and storage automatically. The change allows models to judge if they have enough evidence and perform additional searches if needed.
The agentic retrieval process is exposed through the AgenticRetrieveStream API, which streams trace events back to users as the model plans its steps. In contrast, the standard Retrieve API runs a single hybrid search and returns scored chunks. Both APIs are accessible via the langchain-aws package, allowing developers to choose between the two methods in their LangChain applications.
Amazon Bedrock Managed Knowledge Bases removes the need for self-managed vector stores, embeddings, and re-ranking models, simplifying the RAG architecture. Users can configure a data source, such as an Amazon S3 bucket, and let the service handle the rest. The solution includes two APIs, with the AgenticRetrieveStream API supporting multi-step planning loops for complex queries.
"Agentic retrieval breaks the question into sub-queries, runs them, judges whether it has enough evidence, and searches again if it doesn’t," said Amazon. This approach ensures that the retrieved chunks cover more of the question's intent, leading to more comprehensive answers. The feature is part of Amazon's broader efforts to enhance its machine learning capabilities.
The announcement follows Amazon's expansion of its Amazon Bedrock platform, which includes a range of foundation models and tools for developers.
Amazon did not specify when agentic retrieval will be available beyond the current release, and it remains unclear how it will impact existing applications. The feature is expected to improve the accuracy of RAG applications by enabling more thorough searches.
Source: awsml
