Amazon Bedrock Knowledge Bases: how they work and what they cost

An Amazon Bedrock knowledge base is AWS's managed retrieval-augmented generationRAGA way to answer from your own documents. A search finds the passages that match the question, and they are added to the prompt before the model writes.: it reads your documents, splits and embeds them, keeps a search index and returns the passages that answer a question. A managed knowledge base costs $5 per GB of raw data a month plus $1 per 1,000 searches, with parsing, embeddings and reranking included; the model that writes the answer is billed on top.

It saves you building the pipeline in our RAG app guide by hand. Whether it is the right choice depends on 3 things: where your documents live, how much control you want over the index, and how many questions it will answer.

How a knowledge base works

1. Connect a source. S3, SharePoint, Confluence, Google Drive, OneDrive, a web crawler, or a custom source you push documents into.
2. Sync. Bedrock parses each file (text, PDFs, tables, images), splits it into chunks, about 300 tokens each by default on sentence boundaries, embeds each chunk and writes it to the index. Later syncs are incremental: only added, changed or deleted files are processed.
3. Query. A question is embedded, the closest chunks are found and reranked, and the best 5 (by default) come back with their source and metadata.

The sync is not automatic. If a policy changes in SharePoint and nobody runs a sync, the knowledge base keeps answering from the old one. Start a sync from your upload pipeline or on a schedule.

Managed or customer-managed

There are 2 kinds, and AWS recommends the first:

Choose customer-managed when you already run a Vector DatabaseVector DatabaseA database built to store embeddings and quickly find the ones closest to a query, which is how AI apps search by meaning. you want to keep, need Hybrid SearchHybrid SearchRunning a keyword search (usually BM25) and a vector search for the same query and merging the 2 ranked lists, so exact terms like a product code and loose wording like "money back" are both found. tuned your way, or want to read the index directly. Choose managed for everything else: no store to size, no embedding bill, and the connectors your company documents already sit behind.

The 4 ways to ask

• Retrieve returns the matching chunks with scores and source locations. You build the prompt and call the model yourself. Use it when you want control over the prompt, or when the answer goes to an agent.
• RetrieveAndGenerate retrieves, passes the chunks to the model you name and returns an answer with citations. The quickest way to a working assistant.
• Agentic retrieval (managed only) lets a model break a hard question into several searches and combine them. Better on "compare our 2025 and 2026 travel policies"; costs more per question.
• GenerateQuery turns a question into a query for a structured data store, such as a data warehouse, when the knowledge base is connected to one.

One thing to know: GuardrailsGuardrailChecks placed around a model that stop or change what goes in and what comes out, such as a filter on user messages or an approval step before an agent acts. check the question and the generated answer, not the retrieved passages. If your documents could contain instructions planted by someone else, a classic Prompt InjectionPrompt InjectionAn attack that hides instructions in text a model reads, such as a web page, an email or a file, so the model follows the attacker instead of you. route, treat them as untrusted text.

Metadata filters

Put a <file>.metadata.json file beside each document:

{ "metadataAttributes": { "department": "hr", "year": 2026, "region": ["eu", "uk"] } }

Then filter at query time, so an HR question only searches HR documents from 2026:

"filter": { "andAll": [
  { "equals": { "key": "department", "value": "hr" } },
  { "greaterThanOrEquals": { "key": "year", "value": 2026 } },
  { "listContains": { "key": "region", "value": "eu" } }
] }

Filters are the cheapest accuracy gain there is: fewer wrong candidates means fewer wrong answers. startsWith and stringContains do not work on managed knowledge bases or S3 vector buckets, so store a value you can match exactly.

Managed knowledge bases can also filter by the user's own permissions in SharePoint, OneDrive, Google Drive and Confluence, matched by email address. AWS says plainly this is not an authorization mechanism: it narrows results; your application still decides who may ask.

A worked month

A help desk over 20 GB of policies and manuals, answering 5,000 questions a day (150,000 a month), on a managed knowledge base with RetrieveAndGenerate:

• Index: 20 GB × $5 = $100
• Searches: 150,000 ÷ 1,000 × $1 = $150
• The answer on Claude Sonnet 5.5, 2,000 tokens in (5 chunks plus the question and instructions) and 300 out: 150,000 × (2,000 × $2 + 300 × $10) ÷ 1 million = $1,050

$1,300 a month, and the model is 80% of it. The same answers on Claude Haiku 5.5 cost about $53, so the month drops to about $300. Switch to agentic retrieval at 3 searches a question and the knowledge base part goes from $250 to $1,150: keep it for the questions that need it.

Try itWhat a managed knowledge base costs a month

Raw data indexed

20 GB

Questions a day

5,000

RetrieveAgentic

Index storage 20 GB × $5$100

Retrieve calls 150,000 × $0.001$150

A month: $250$1 per 1,000 searches

What to do

1. Start with a managed knowledge base and the default chunking; change chunk size only when tests show wrong passages, as in our chunk size test.
2. Add metadata from day one. Adding it later means re-syncing everything.
3. Sync from your pipeline, not by hand.
4. Measure with a small set of real questions and expected sources before you tune anything.
5. Spend on the model, not the search: the search is a dollar per thousand; the answer costs far more. Our Bedrock pricing guide covers how to cut the model bill, and the embedding cost calculator prices a customer-managed index.