AWS has moved Amazon Bedrock Managed Knowledge Base into general availability, positioning it as a managed retrieval layer for enterprise AI agents and retrieval-augmented generation workloads. The launch matters because it targets one of the messiest parts of shipping enterprise AI systems: connecting scattered internal data, parsing mixed document formats, enforcing permissions, and getting reliable retrieval without forcing teams to assemble their own vector, graph, and orchestration stack.

According to the AWS Machine Learning Blog, the service is designed to let developers create a knowledge base, attach data sources, and begin ingestion with minimal setup, while retaining deeper controls for teams that want to tune embeddings, rerankers, and chunking later. AWS says the goal is to reduce a process that often takes days or weeks down to minutes through defaults and managed infrastructure. That is a vendor claim, but it points to a real market pressure point as enterprises try to move from AI demos to production systems with grounded answers and policy-aware access.

What AWS is launching

At the center of the release is Amazon Bedrock Managed Knowledge Base, which AWS describes as a fully managed agentic retrieval service. In practical terms, it combines ingestion, parsing, storage, retrieval, and access controls into one AWS-managed layer inside Amazon Bedrock.

The company says teams no longer need to separately provision a vector database, decide similarity metrics, manage scaling, or stitch connectors and retrieval infrastructure together. Instead, the service handles ingestion from supported sources, manages the underlying storage automatically, and exposes retrieval APIs for both straightforward search and more complex multi-step retrieval.

AWS is pitching this as infrastructure for enterprise search, internal copilots, and agentic RAG systems. That framing is important. The announcement is not just about a new search feature; it is AWS trying to make Amazon Bedrock more attractive as an end-to-end runtime for enterprise AI applications, especially for teams that want agents to pull from company documents safely.

The native connector list is a notable part of the launch. AWS says Managed Knowledge Base currently includes connectors for Amazon S3, Microsoft SharePoint, Atlassian Confluence, Google Drive, Microsoft OneDrive, and a Web Crawler, plus a direct ingestion API for unsupported sources. Those integrations cover common enterprise knowledge silos, which is often where early deployment friction shows up.

Why retrieval remains the hard part of enterprise AI

AWS’s blog post is useful because it identifies the operational burden many teams already know firsthand. Building a grounded enterprise AI system usually means choosing ingestion tools, document parsers, storage engines, embedding models, chunking strategies, retrieval logic, and then layering on observability and security. Each choice creates another integration surface and another operational dependency.

Managed Knowledge Base is AWS’s attempt to collapse those decisions into a default path. The company says users do not need to choose a model to get started in the console and can defer many tuning decisions until later. For product teams under pressure to launch internal search or question-answering tools quickly, that may be the most valuable part of the release.

AWS also emphasizes mixed-format parsing. According to the company, the service can handle visual documents such as PDFs, PPT and PPTX, and DOCX files up to 500 MB, as well as audio files up to 2 GB and video files up to 10 GB. It says the system automatically selects parsing strategies for tables, charts, diagrams, mixed layouts, and media. If that works well in production, it could spare builders from maintaining separate preprocessing pipelines for different enterprise content types.

The company’s storage abstraction is another core bet. Rather than exposing direct management of an underlying vector or graph layer, AWS says it auto-provisions and auto-scales a unified storage layer and keeps hybrid search, meaning keyword plus semantic retrieval, continuously enabled. That may appeal to enterprise teams that care more about answer quality and governance than about tuning database internals.

Two retrieval paths for simple search and agent workflows

AWS is separating the product into two main retrieval patterns. The first is a standard Retrieve API, which returns ranked source chunks with metadata and relevance scores. AWS says this is intended for direct lookups, FAQ-style interactions, and other lower-latency scenarios.

The second is more strategically significant: Agentic Retrieval. AWS says this mode uses a foundation model to break a complex question into sub-queries, search across one or more knowledge bases, assess whether the results are sufficient, and if needed run additional retrieval rounds. According to the company, it can also synthesize a final response using either a managed orchestration model or another model available through Amazon Bedrock.

That design aligns with a broader market shift. Enterprises increasingly want AI agents to do more than fetch a single passage. They want systems that can compare policies, synthesize findings from multiple documents, and follow multi-hop reasoning chains across repositories. If AWS can make that orchestration reliable enough, it gives Amazon Bedrock a stronger claim as a practical agent platform rather than only a model access layer.

Still, “agentic retrieval” should be read carefully. AWS describes a loop of planning, retrieving, and evaluating, with up to five retrieval rounds by default. That suggests more capable retrieval on paper, but also more moving parts that can introduce latency, cost, and failure modes. The company did not provide independent benchmarks in the cited material comparing retrieval quality, latency, or cost against alternative stacks.

Security and enterprise controls are central to the pitch

The strongest enterprise angle in the launch is security and permissions handling. AWS says Managed Knowledge Base uses real-time ACL checks in addition to pre-retrieval ACL filtering. The pre-filtered documents, according to the company, are transient for the duration of the API call and are not visible to large language models or users.

That architecture matters because stale permission mappings are a common problem in enterprise search and RAG systems. If a retrieval layer indexes content without reflecting current source permissions, employees can end up seeing material they should not access. AWS says its query-time checks rely on the authoritative source rather than potentially outdated copied ACL data.

Encryption is also part of the message. AWS says data is encrypted in transit and at rest using AWS KMS keys, either AWS-managed or customer-managed. That will not eliminate buyer concerns around data residency, auditability, and model behavior, but it does line up with procurement requirements for enterprise AI deployments already centered on AWS.

For enterprise buyers, this may be the real buying argument: less effort to get grounded retrieval running, with access controls enforced close to the source systems they already trust.

Evidence, customer claims, and what is still unverified

The evidence in this story comes almost entirely from AWS-controlled sources, specifically AWS coverage and the AWS Machine Learning Blog. That means the most favorable claims about setup time, retrieval quality, scale, and customer use should be treated as vendor-reported unless independently confirmed.

AWS included customer statements from Syngenta Group and MRH Trowe. Syngenta Group’s cited executive said the company uses Bedrock Managed Knowledge Bases to let employees create knowledge bases on demand using SharePoint and Confluence data for internal search and agentic RAG applications. MRH Trowe said it is using the product for an internal AI copilot spanning thousands of documents in Confluence and SharePoint across English and German content.

AWS also included a statement attributed to OpenAI saying it is using Bedrock Managed Knowledge Bases’ RAG capabilities to ground inference and model responses at scale for millions of users with the right customer context. That is the most eye-catching adoption signal in the announcement, but the source material does not provide deployment details, scope, timing, or independent verification. As presented, it is a vendor-published customer quote rather than a reported partnership profile.

What is missing from the launch materials is also notable. AWS did not include third-party benchmark data, side-by-side comparisons against self-managed retrieval stacks, or clear pricing examples in the cited evidence. For builders evaluating Amazon Bedrock against rival options, those omissions leave open questions about total cost, tuning flexibility, and real-world retrieval quality under enterprise workloads.

What this means for builders and enterprise teams

For AI builders, the release could remove a large amount of infrastructure work from early product development. Teams building internal copilots, workflow assistants, or AI agents often spend more time on ingestion and permissions than on prompts or app logic. A managed path inside Amazon Bedrock may let them prototype faster and keep more of the architecture under one cloud contract.

For enterprise architects, the tradeoff is familiar. Managed services can reduce operational burden, but they also abstract away implementation details that advanced teams may want to control directly. Some companies will welcome not having to choose or tune a vector store. Others may prefer explicit control over retrieval databases, indexing, chunking pipelines, and reranking stacks, especially if they are already using custom RAG infrastructure.

The connector support for Microsoft SharePoint, Atlassian Confluence, Google Drive, Microsoft OneDrive, and Amazon S3 makes the product immediately relevant to internal knowledge search projects. The Web Crawler option broadens it beyond private repositories, which could be useful for agents that need a mix of internal and public information. But buyers will still need to test how well the service handles noisy documents, multilingual corpora, and domain-specific terminology.

In market terms, AWS is strengthening the argument that enterprise AI infrastructure should look more like a managed application platform and less like a collection of point tools. The more capabilities Amazon Bedrock absorbs, the harder it becomes for standalone retrieval vendors to compete purely on convenience. That does not guarantee AWS will win on quality or price, but it increases pressure on the broader RAG ecosystem.

What to watch next

The next signals to watch are practical rather than rhetorical. First, pricing clarity: does Managed Knowledge Base lower all-in deployment costs once ingestion, retrieval, and agent orchestration are included? Second, retrieval quality: AWS will need reference architectures, benchmark data, or independent case studies that show how Agentic Retrieval performs on real enterprise tasks.

Third, connector expansion will matter. The current list is useful, but enterprise data rarely lives in only six places. Fourth, governance features will likely become a key battleground, including audit trails, policy controls, and debugging tools for failed retrieval chains. And finally, watch whether AWS turns the service into a broader standard layer for AI agents across Amazon Bedrock rather than a standalone retrieval feature.

Creati.ai perspective

This launch addresses a real bottleneck in enterprise AI: retrieval infrastructure is still too fragmented for most product teams to operationalize quickly. By packaging ingestion, permissions, parsing, storage, and multi-step retrieval into Amazon Bedrock, AWS is trying to make grounded enterprise AI deployable by default rather than by custom integration.

The open question is whether convenience will translate into dependable performance. Enterprise buyers will not adopt a managed retrieval layer on setup simplicity alone. They will want proof that Amazon Bedrock can deliver consistent relevance, correct access enforcement, and acceptable latency under messy real-world conditions. If AWS can show that, Managed Knowledge Base could become one of the more consequential pieces of infrastructure for AI agents inside large organizations.