Amazon Bedrock Enhances AI Knowledge Retrieval with Single-Shot and Agentic Modes

Amazon Web Services (AWS) has expanded its Amazon Bedrock Managed Knowledge Base with new retrieval options that adapt AI response efforts based on query complexity. The update supports three modes—single-shot retrieval augmented generation (RAG), agentic multi-step retrieval, and retrieval-only—through the AgenticRetrieveStream API.

Single-shot RAG performs a single search of the knowledge base and generates an answer from the retrieved passages, ideal for straightforward inquiries. Agentic retrieval breaks down complex questions into sub-queries, iteratively searching multiple knowledge bases and refining results before answering. Retrieval-only mode returns relevant source passages without generating an AI-produced answer, allowing users to review text directly.

These modes are configurable via the maxAgentIteration parameter, which controls the number of planning and retrieval rounds. Setting maxAgentIteration to 1 triggers single-shot RAG, 0 returns passages without answer generation, and values from 2 to 5 enable iterative agentic retrieval with multiple searches. This flexibility lets applications balance response speed, cost, and thoroughness on a per-query basis without maintaining separate systems.

The banking sector example illustrates practical benefits: an internal AI assistant uses single-shot RAG to quickly provide a checklist of documents for opening a business account, while agentic retrieval enables detailed comparisons of account-opening requirements across different business structures, producing answers supported by citations from multiple policy documents. Retrieval-only responses support direct review of source material where summaries are insufficient.

From a developer perspective, these capabilities simplify application design by consolidating retrieval workflows into a single API, diminishing the need for custom query planning. Additionally, single-shot RAG and retrieval-only modes reduce latency and potential costs by minimizing foundation model calls, enhancing the system's efficiency for routine tasks.

While the agentic mode offers improved accuracy for multifaceted queries, it increases latency due to multiple retrieval iterations and model invocations. Conversely, single-shot RAG may be less comprehensive on complex questions but offers faster results suitable for well-scoped queries. Retrieval-only mode lacks AI-generated synthesis, which can complicate user interpretation but preserves access to original content.

Businesses adopting these options are advised to benchmark with representative questions and documents to determine the optimal maxAgentIteration setting, balancing speed and accuracy according to their workflows. Amazon Bedrock's integration of these modes allows seamless adjustment and incremental adoption, supporting diverse information retrieval needs within a unified system.