--- title: "Amazon Bedrock Knowledge Bases — managed RAG, four chunking strategies, and the Retrieve vs RetrieveAndGenerate split" date: 2026-06-06 service: "Amazon Bedrock" component: "Knowledge Bases" tags: [bedrock, knowledge-bases, rag, retrieve, retrieve-and-generate, generate-query, chunking, hierarchical-chunking, semantic-chunking, embeddings, vector-store, opensearch-serverless, s3-vectors, reranking, citations, multimodal] source: https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html verified_on: 2026-06-06 url: https://vanemmerik.ai/aws-ai/2026-06-06.html --- # AWS Bedrock & AgentCore · Tip of the Day · 2026-06-06 ## Amazon Bedrock Knowledge Bases — managed RAG without the plumbing Yesterday was about *measuring* an agent. Today is about *feeding* it. **Amazon Bedrock Knowledge Bases** is the managed Retrieval-Augmented Generation (RAG) service: point it at your data, pick an embedding model and a vector store, and it handles the parse → chunk → embed → index pipeline and the query-time retrieval — so your application gets relevant context (with citations) instead of a model guessing from training data alone. aws bedrock-agent-runtime retrieve-and-generate \ --input '{"text":"What is our refund window?"}' \ --retrieve-and-generate-configuration '{...knowledgeBaseId...}' ≈ 9 min read · Amazon Bedrock · Knowledge Bases ## 01 · What Knowledge Bases actually is From the [overview](https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html): RAG "uses information from data sources to improve the relevancy and accuracy of generated responses." Knowledge Bases is AWS's out-of-the-box implementation — it "abstracts from the heavy lifting of building pipelines" so you don't hand-roll the ingestion and retrieval layer, and it removes "the need to continually train your model to be able to use your private data." Setting one up is four steps: 1. (Optional) Stand up a supported vector store — or let the console create an Amazon OpenSearch Serverless store for you. 2. Connect the knowledge base to an unstructured or structured data source. 3. Sync the data source so it's ingested into the index. 4. Query it from your application or agent — return raw sources, generate a natural-language answer, or transform the question into a structured query (e.g. SQL). ## 02 · The ingestion pipeline: parse → chunk → embed → index For unstructured data, ingestion converts each document into text, splits it into chunks, converts each chunk to a vector embedding, and writes those embeddings to a vector index "while maintaining a mapping to the original document." Those vectors are what make semantic search possible: at query time the user's question is embedded too, and the index returns the chunks whose vectors sit closest to it. Before chunking, you can choose a **parser**. The default parser reads text; for documents heavy with tables, figures, or scanned pages you can use a foundation-model parser — the docs list **Claude vision**, **Nova vision**, and **Llama 4 vision** model families — or the **Bedrock Data Automation (BDA) parser** (in preview, US West (Oregon) only, subject to change). ## 03 · Four chunking strategies Chunking is the single most consequential ingestion choice. Bedrock supports four text strategies: - **Fixed-size** — you set a maximum tokens-per-chunk and an overlap percentage between consecutive chunks. - **Default** — splits into chunks of "approximately 300 tokens," honoring sentence boundaries so complete sentences stay intact. - **No chunking** (`NONE`) — each document becomes a single chunk. Pre-split your files first if you go this route. The catch: you "cannot view page number in citation or filter by the *x-amz-bedrock-kb-document-page-number* metadata field." - **Hierarchical** — nested parent and child chunks. Retrieval pulls the precise child chunks, then "replaces them with broader parent chunks" to hand the model more context. You set parent size, child size, and the overlap token count. - **Semantic** — splits on meaning, not syntax, using three hyperparameters: maximum tokens, **buffer size** (surrounding sentences pulled in for each embedding — a buffer of 1 embeds the previous, current, and next sentence together), and a **breakpoint percentile threshold** (a higher threshold means fewer, larger chunks). For multimodal content the rules differ: with **Nova multimodal embeddings**, chunking happens at the embedding-model level — audio and video chunk duration is configurable from **1–30 seconds** (default 5) — and text chunking strategies apply only to text documents. The BDA parser instead converts audio/video to transcripts and scene summaries first, then applies the text strategies. ## 04 · Where the vectors live Knowledge Bases can index into a range of vector stores. You either let the console spin up an OpenSearch Serverless collection, or bring your own: - **Amazon OpenSearch Serverless** (console can auto-create it) - **Amazon OpenSearch Managed Clusters** - **Amazon S3 Vectors** (cost-optimized vector storage for RAG) - **Amazon Aurora PostgreSQL** (pgvector) - **Amazon Neptune Analytics** (for graph-backed retrieval) - **Pinecone**, **Redis Enterprise Cloud**, **MongoDB Atlas** (credentials brokered through AWS Secrets Manager) ## 05 · Querying: Retrieve vs RetrieveAndGenerate vs GenerateQuery Three runtime APIs, each a level up in how much work AWS does for you: - **`Retrieve`** — returns the source chunks (or images) most relevant to the query as an array. You own the prompt assembly and generation. - **`RetrieveAndGenerate`** — joins `Retrieve` with `InvokeModel` to do the whole RAG loop: retrieve the chunks, generate a natural-language answer, and attach **citations** to the specific source chunks. If the source includes visual elements, the model can use insights from those images and attribute them. - **`GenerateQuery`** — converts a natural-language question into a query suited to a structured data store (e.g. SQL). `RetrieveAndGenerate` is the combined action: under the hood it uses `GenerateQuery` (for structured stores), `Retrieve`, and `InvokeModel`. Because `Retrieve` is exposed on its own, you "have the flexibility to decouple the steps in RAG and customize them." With either retrieval API you can add a **reranking model** to re-order results by relevance before they reach the prompt. ## 06 · Embeddings, multimodal, and structured data The embedding model turns text into the vectors the index compares. Supported models and their vector types: - **Titan Embeddings G1 – Text** — floating-point, 1536 dimensions. - **Titan Text Embeddings V2** — floating-point or **binary**; 256 / 512 / 1024 dimensions. - **Cohere Embed English v3** / **Multilingual v3** — floating-point or binary, 1024 dimensions. - **Titan Multimodal Embeddings G1** and **Cohere Embed v3 (Multimodal)** — 1024 dimensions, for image and text. Binary vectors use 1 bit per dimension instead of 32, so they're far cheaper to store — but less precise, and they require both a model and a vector store that support binary. Beyond plain text, Knowledge Bases can extract and retrieve images from visually rich documents, accept images as queries, convert natural language to SQL against structured stores, build on an **Amazon Kendra GenAI index** or **Neptune Analytics graphs**, and plug into an **Amazon Bedrock Agents** workflow. It also supports **inference profiles** for cross-Region inference to raise throughput on the parsing and generation models. ## Limits worth knowing - **`NONE` chunking drops page-number citations.** No `x-amz-bedrock-kb-document-page-number` filter, no page numbers in citations — pre-split your files if granularity matters. - **Hierarchical chunking isn't recommended with an S3 vector bucket.** Combined parent+child token counts over ~8,000 can hit metadata size limits. Hierarchical retrieval can also return *fewer* results than requested, since child chunks get replaced by parents. - **Semantic chunking costs extra.** It invokes a foundation model during ingestion, billed on top of embedding cost. - **Cross-Region inference shares data across Regions.** The docs flag this explicitly — factor it into your data-residency rules before enabling inference profiles. - **Custom and SageMaker models need explicit prompts.** If you bring your own generation model, you must supply the orchestration and generation prompt templates with the required input variables. - **The BDA parser is preview, Oregon-only.** Treat it as subject to change. ## Try it in five minutes 1. In the Bedrock console, create a knowledge base; let it create an **OpenSearch Serverless** store so you skip the vector-store setup. 2. Point a data source at an S3 prefix of PDFs or Markdown. Start with **default chunking** (≈300 tokens) — tune later. 3. Pick **Titan Text Embeddings V2** as the embedding model and **Sync**. Wait for ingestion to finish. 4. Open **Test knowledge base**, toggle *Generate responses* on, ask a question, and inspect the **citations** — every claim should trace to a source chunk. 5. When you need control, drop to the API: call `Retrieve` for raw chunks, or `RetrieveAndGenerate` for an answer-with-citations in one request. Add a reranker if the top results look noisy. Once that loop feels natural, swap default chunking for **hierarchical** or **semantic** and re-sync to see how retrieval quality shifts on your own corpus. ## Verified against the official AWS docs on 2026-06-06 Sources: - Retrieve data and generate AI responses with Amazon Bedrock Knowledge Bases — https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base.html - How Amazon Bedrock knowledge bases work — https://docs.aws.amazon.com/bedrock/latest/userguide/kb-how-it-works.html - Retrieving information from data sources — https://docs.aws.amazon.com/bedrock/latest/userguide/kb-how-retrieval.html - How content chunking works — https://docs.aws.amazon.com/bedrock/latest/userguide/kb-chunking-parsing.html - Supported models and Regions — https://docs.aws.amazon.com/bedrock/latest/userguide/knowledge-base-supported.html If the docs change, this tip is a snapshot of that day — check the sources for current behaviour. --- *This page — research, writing, verification, and deployment — was built by Claude Cowork. The tip was generated this morning, cross-checked against the official AWS docs, and published to Cloudflare R2 on a schedule. A daily experiment by Monty van Emmerik · vanemmerik.ai*