Amazon Bedrock Knowledge Bases: managed RAG, and which of the two types to pick.

IWhat a Bedrock Knowledge Base is

A model only knows its training data. A Knowledge Base gives it your data at question time, with citations.

Retrieval-augmented generation (RAG) finds the passages of your documents that match a question and puts them into the model's prompt, so the answer comes from your facts. It is cheaper to update and easier to audit than fine-tuning a model on the same data.

Building RAG by hand means running a loader, a parser, a chunker, an embedding model, a vector store and the query loop, and keeping all of it in sync with the source. A Knowledge Base turns that into one service: you connect a data source, run a sync, and call an API.

Two choices are still yours: which type of knowledge base to build (next section), and, on a customer-managed one, the parsing, chunking, embedding model and vector store. Those choices drive answer quality and cost.

IIBedrock Managed vs customer-managed knowledge bases

AWS added a fully managed type. It is now the one AWS recommends for most builds.

With a Bedrock Managed Knowledge Base, Bedrock runs ingestion, storage, indexing and retrieval. You provide the data sources. Embedding and reranking use service-managed models by default, and you can bring your own Bedrock embedding model (float32, 1,024 dimensions) or reranker. With a customer-managed knowledge base you choose, provision and scale the vector store yourself, and you have direct access to it.

IIIData sources you can connect

Which connectors you get depends on the knowledge base type.

• Amazon S3: Both types. Put PDFs, text, HTML, Markdown, Word or CSV files in a bucket and point the knowledge base at the prefix. Add a metadata file per document if you want to filter on it later. Images, audio and video are supported only from S3 and custom sources.
• Custom data source: Both types. Push documents through the API from any system, useful when you already have a pipeline.
• SharePoint, Confluence, Web Crawler: Native connectors on Bedrock Managed knowledge bases. Since 30 September 2026 new connectors of these types (and Salesforce) cannot be created on customer-managed knowledge bases; existing ones keep working.
• Google Drive and OneDrive: Native connectors on Bedrock Managed knowledge bases.

One knowledge base can have several data sources. Each source syncs on demand or on a schedule, and the next sync picks up added, changed and deleted documents. Scope each connector to the spaces, sites or prefixes you need; managed knowledge bases also offer access control list awareness. See the connector page for current details.

IVParsing and chunking

How documents are parsed and cut into chunks matters more for answer quality than the model you pick.

VEmbedding models and vector stores (customer-managed)

On a customer-managed knowledge base you choose both. On a managed one, Bedrock handles them.

Common Bedrock embedding models are Amazon Titan Text Embeddings, Cohere Embed (English and multilingual) and Amazon Nova Multimodal Embeddings. Vectors from one model cannot be compared with vectors from another, so changing the embedding model means re-embedding the whole corpus. Choose once, on purpose.

The vector store is the piece you see and pay for directly. On 10 October 2026 the setup page listed eight options. Bedrock can quick-create an index for some of them.

VIRetrieve, RetrieveAndGenerate, agentic retrieval and metadata filters

The API you call decides how much of the RAG loop Bedrock runs for you.

• Retrieve: Returns the top matching chunks with scores and source locations, without calling a model. Use it when you want your own prompt, your own model choice, reranking, or an agent in the loop.
• RetrieveAndGenerate: Retrieves, builds the prompt, calls the model you name and returns an answer with citations, with session support for follow-up questions. The fastest way to ship "chat with your docs".
• Agentic retrieval: Managed knowledge bases only. A model plans the retrieval steps for harder questions. Priced per agentic call plus the underlying Retrieve calls.
• Metadata filters: Tag documents with fields such as department, date or tenant, then filter at query time. This sharpens results and is the main building block for multi-tenant access control. Pair it with Bedrock Guardrails for content safety.

VIIKnowledge Bases vs building RAG yourself

Managed wins on time and upkeep. DIY wins on control.

A Knowledge Base gets you a synced, searchable corpus with citations in hours, and AWS runs the pipeline. Building your own RAG (loaders, a framework such as LangChain or LlamaIndex, an embedding call, a vector store) lets you use any embedding model, any search method and any chunking logic. A middle path is common: use a knowledge base with the Retrieve API for ingestion and search, and own the prompt and model call yourself.

Start managed if your needs are standard: index documents, retrieve passages, answer with citations. Build your own only when you hit a specific wall, such as an embedding model Bedrock does not offer or a retrieval method you must control. See RAG on AWS for the wider architecture.

VIIIWhat a Knowledge Base costs

The two types are billed differently.

Bedrock Managed. On 10 October 2026 the pricing page listed index storage at $5.00 per GB of raw data per month, standard retrieval at $1.00 per 1,000 calls, and agentic retrieval at $4.00 per 1,000 agentic calls plus $1.00 per 1,000 underlying Retrieve calls. Parsing, embeddings and reranking with the managed components were listed as included. Bringing your own embedding or reranking model adds charges. RetrieveAndGenerate also bills the answering model's tokens.

Customer-managed. You pay the embedding model per token (at ingest and for every query), the vector store you run (often the biggest standing cost, because some stores charge for capacity even when idle), model parsing per page if you use it, semantic chunking's model use, reranking per query, and the answering model's tokens.

Two patterns hold for both. Retrieved passages become input tokens on every answer, so returning fewer, better chunks cuts cost as well as noise. And caching a stable system prompt with prompt caching lowers the repeated part of each call.

IXWho can file AWS funding for a RAG build

There are two kinds of AWS money for a Bedrock build. A partner files one. You apply for the other yourself.

Partner-filed funding. An AWS consulting partner can file POC (proof of concept) funding to test a Bedrock workload such as a RAG assistant, whether or not you run on AWS today. If you already run on AWS and are adding new AI work, the partner can file incremental workload (IW) funding instead. For startups this is often $50K to $150K. Depends on the workload; AWS approves. The work has to be real, and the partner has to be the one doing it.

AWS Activate, which you apply for yourself. Activate Founders gives $1,000 to start, up to $5,000, with no investor needed. Activate Portfolio gives up to $200,000 if your VC, accelerator or angel group is an Activate Provider and gives you its Org ID. A consulting partner cannot apply for Activate on your behalf. AWS made Activate credits usable for third-party models on Bedrock in April 2024.

For most startups without a VC, the partner route is the larger of the two. CloudRoute matches you with one vetted AWS partner for free. Partners pay CloudRoute a monthly membership plus 10% of AWS-approved funding.