Amazon Bedrock Knowledge Base¶

Supported in ADKPython

Connect your ADK agents to Amazon Bedrock Managed Knowledge Bases for retrieval-augmented generation (RAG). Agents can search enterprise documents and get grounded answers without hallucinating.

Why Bedrock Knowledge Bases for ADK?¶

Amazon Bedrock Managed Knowledge Bases provide fully managed RAG infrastructure — no vector store to provision, no embeddings to configure, automatic scaling. Combined with ADK agents, your agent can:

Search enterprise documents (PDFs, web pages, databases) in natural language
Use agentic retrieval for multi-hop reasoning across documents
Get grounded answers with source citations
Scale automatically without managing infrastructure

Use cases¶

Enterprise Q&A agents: Answer questions from internal documentation, policies, and knowledge bases
Customer support agents: Retrieve relevant help articles and product information
Research assistants: Search across large document collections with multi-hop reasoning
Compliance agents: Look up regulatory documents and provide cited answers

Prerequisites¶

AWS account with Amazon Bedrock access
A Bedrock Managed Knowledge Base created (via AWS Console or Terraform)
AWS credentials configured (environment variables, IAM role, or AWS profile)
Python 3.10+

Installation¶

pip install google-adk-community boto3>=1.43.2

Use with agent¶

from google.adk.agents import Agent
from google.adk_community.tools.bedrock_kb import bedrock_kb_retrieve

# Create an agent with Bedrock KB retrieval
agent = Agent(
    model="gemini-2.0-flash",
    tools=[bedrock_kb_retrieve],
    instruction="You are a helpful assistant. Use bedrock_kb_retrieve to search the knowledge base when answering questions about company policies or documentation.",
)

The agent will automatically call bedrock_kb_retrieve when it needs information from the knowledge base.

Configuration via environment variables¶

export KNOWLEDGE_BASE_ID="YOUR_KB_ID"
export AWS_REGION="us-west-2"
export USE_AGENTIC_RETRIEVAL="true"  # Multi-hop reasoning (default)

Configuration via function arguments¶

# Pass knowledge_base_id directly (overrides env var)
result = bedrock_kb_retrieve(
    query="What is our refund policy?",
    knowledge_base_id="YOUR_KB_ID",
    max_results=5,
)

Available tools¶

Tool	Description
`bedrock_kb_retrieve`	Searches a Bedrock Managed Knowledge Base and returns relevant passages with sources and scores. Supports agentic retrieval (multi-hop reasoning) with automatic fallback to standard search.

Tool parameters¶

Parameter	Type	Description
`query`	`str`	The natural language question or search query (required)
`knowledge_base_id`	`str`	Bedrock KB ID. Defaults to `KNOWLEDGE_BASE_ID` env var
`max_results`	`int`	Maximum results to return. Defaults to 5

Retrieval modes¶

Agentic retrieval (default)¶

Uses AgenticRetrieveStream — the model reasons over multiple retrieval passes, decomposes complex queries, and applies managed reranking for better results.

export USE_AGENTIC_RETRIEVAL="true"

Standard retrieval¶

Uses Retrieve with managedSearchConfiguration — single-pass semantic search.

export USE_AGENTIC_RETRIEVAL="false"

IAM permissions¶

Your AWS credentials need:

{
    "Effect": "Allow",
    "Action": [
        "bedrock:Retrieve",
        "bedrock:AgenticRetrieveStream"
    ],
    "Resource": "arn:aws:bedrock:REGION:ACCOUNT:knowledge-base/KB_ID"
}