AWS has added two retrieval modes to its knowledge base. The central judgment is that enterprise AI is moving away from sending every question through the same expensive process toward allocating cost and time according to question difficulty.
What this is
Amazon Bedrock Managed Knowledge Base now supports single-shot Retrieval Augmented Generation (RAG, short for the process of first finding supporting evidence in enterprise materials and then having the model answer), a retrieval-only mode that returns matching source text, and the existing agentic retrieval.
The differences are concrete. Single-shot RAG searches the materials once and generates an answer, making it suitable for checking lists. Retrieval-only mode provides raw excerpts, allowing applications or employees to verify them. Agentic retrieval breaks a question into subquestions, searches repeatedly, continues gathering information when the available material is insufficient, and then answers with citations. AWS illustrates this with a bank account-opening policy assistant: checking a single requirement does not require multiple rounds of search, while comparing sole proprietorships, limited liability companies, and partnerships requires evidence gathering across multiple documents.
Industry view
We are noticing that AWS is pushing cost control in large language model applications from the model layer down to the retrieval layer. In the past, many teams enabled multi-round search and automatic answer generation by default to avoid missing information. The new modes allow developers to make more granular trade-offs among latency, cost, and answer depth.
But this capability does not become reliable merely by connecting it. Multi-round retrieval may expand the scope of data access and depends more heavily on permissions, document quality, and stopping conditions. If a model decomposes a question incorrectly or retrieves an outdated policy, the answer may still be wrong even with citations, now packaged with greater persuasive force.
Another risk is convergence at the interface level and in functionality. After AWS offers simple, agentic, and retrieval-only paths simultaneously, application teams still need to establish evaluation mechanisms that separately test accuracy, citation hit rate, response time, and cost per request. Otherwise, so-called on-demand configuration can easily degenerate into defaulting to the most complex and expensive mode.
Impact on regular people
For enterprise IT: Knowledge base development no longer has to focus on a single “strongest” workflow. Simple inquiries can use a low-cost path, while complex verification can invoke multi-round search. This can also help departments attribute AI usage costs.
For professionals: Internal assistants will offer more control over response speed and the scope of source material. However, when questions involve approvals, accounts, or compliance decisions, users should still check the original documents; citation links cannot replace professional review.
For the consumer market: High-frequency use cases such as customer service and membership inquiries will prioritize faster, cheaper question-and-answer experiences. At the same time, giving enterprises a “view the original text only” entry point makes it easier for consumers to examine the evidence behind an AI answer.