I implemented a simple RAG by integrating Amazon Bedrock Knowledge Bases (S3 Vectors) with Amazon Bedrock AgentCore

{
    "Version": "2012-10-17",
    "Statement": [
        {
            "Sid": "ECRImageAccess",
            "Effect": "Allow",
            "Action": [
                "ecr:BatchGetImage",
                "ecr:GetDownloadUrlForLayer"
            ],
            "Resource": [
                "arn:aws:ecr:us-west-2:123456789012:repository/*"
            ]
        },
        {
            "Effect": "Allow",
            "Action": [
                "logs:DescribeLogStreams",
                "logs:CreateLogGroup"
            ],
            "Resource": [
                "arn:aws:logs:us-west-2:123456789012:log-group:/aws/bedrock-agentcore/runtimes/*"
            ]
        },
        {
            "Effect": "Allow",
            "Action": [
                "logs:DescribeLogGroups"
            ],
            "Resource": [
                "arn:aws:logs:us-west-2:123456789012:log-group:*"
            ]
        },
        {
            "Effect": "Allow",
            "Action": [
                "logs:CreateLogStream",
                "logs:PutLogEvents"
            ],
            "Resource": [
                "arn:aws:logs:us-west-2:123456789012:log-group:/aws/bedrock-agentcore/runtimes/*:log-stream:*"
            ]
        },
        {
            "Sid": "ECRTokenAccess",
            "Effect": "Allow",
            "Action": [
                "ecr:GetAuthorizationToken"
            ],
            "Resource": "*"
        },
        {
            "Effect": "Allow",
            "Action": [
                "xray:PutTraceSegments",
                "xray:PutTelemetryRecords",
                "xray:GetSamplingRules",
                "xray:GetSamplingTargets"
            ],
            "Resource": [
                "*"
            ]
        },
        {
            "Effect": "Allow",
            "Resource": "*",
            "Action": "cloudwatch:PutMetricData",
            "Condition": {
                "StringEquals": {
                    "cloudwatch:namespace": "bedrock-agentcore"
                }
            }
        },
        {
            "Sid": "GetAgentAccessToken",
            "Effect": "Allow",
            "Action": [
                "bedrock-agentcore:GetWorkloadAccessToken",
                "bedrock-agentcore:GetWorkloadAccessTokenForJWT",
                "bedrock-agentcore:GetWorkloadAccessTokenForUserId"
            ],
            "Resource": [
                "arn:aws:bedrock-agentcore:us-west-2:123456789012:workload-identity-directory/default",
                "arn:aws:bedrock-agentcore:us-west-2:123456789012:workload-identity-directory/default/workload-identity/agent-*"
            ]
        },
        {
            "Sid": "BedrockModelInvocation",
            "Effect": "Allow",
            "Action": [
                "bedrock:InvokeModel",
                "bedrock:InvokeModelWithResponseStream",
                "bedrock:ApplyGuardrail"
            ],
            "Resource": [
                "arn:aws:bedrock:*::foundation-model/*",
                "arn:aws:bedrock:us-west-2:123456789012:*"
            ]
        },
        {
            "Sid": "KBManagement",
            "Effect": "Allow",
            "Action": [
                "bedrock:CreateKnowledgeBase",
                "bedrock:GetKnowledgeBase",
                "bedrock:UpdateKnowledgeBase",
                "bedrock:DeleteKnowledgeBase",
                "bedrock:ListKnowledgeBases",
                "bedrock:TagResource",
                "bedrock:UntagResource"
            ],
            "Resource": "*"
        },
        {
            "Sid": "KBDataSourceManagement",
            "Effect": "Allow",
            "Action": [
                "bedrock:CreateDataSource",
                "bedrock:GetDataSource",
                "bedrock:UpdateDataSource",
                "bedrock:DeleteDataSource",
                "bedrock:ListDataSources",
                "bedrock:StartIngestionJob",
                "bedrock:GetIngestionJob",
                "bedrock:ListIngestionJobs",
                "bedrock:Retrieve"
            ],
            "Resource": "*"
        }
    ]
}

import os
import logging
from strands import Agent, tool
from strands.models import BedrockModel
from bedrock_agentcore.runtime import BedrockAgentCoreApp
from strands_tools import retrieve as retrieve_module

LOG_LEVEL = os.getenv("LOG_LEVEL", "INFO").upper()
logging.basicConfig(
    level=getattr(logging, LOG_LEVEL, logging.INFO),
    format="%(asctime)s %(levelname)s %(name)s - %(message)s",
)
logger = logging.getLogger("strands")

app = BedrockAgentCoreApp()

SYSTEM_PROMPT = """
# RAG System Prompt

## Basic Policy
You are an assistant that provides accurate and useful answers based on search results. Please follow the guidelines below when responding.

## Response Rules

### 1. Using Search Results
- Use only the provided search results as your information source
- For information not included in the search results, do not speculate or supplement with general knowledge; clearly state "not included in the search results"
- When there are multiple search results, integrate them to create a comprehensive response

### 2. Citation Method
- Always include citation numbers for information in your response (e.g., [1], [2])
- Include a "## References" section at the end of your response with the following format:

### 3. Response Format
- Answer in clear and concise Japanese
- Use headings and bullet points to organize important points
- For lengthy responses, include a summary at the beginning

### 4. When Information is Insufficient
- If search results are not sufficient for the question, clearly state this
- If you can partially answer, clearly indicate which aspects you can address

## Response Example

**Question:** Please explain how solar power generation works.

**Answer:**

Solar power generation is a system that directly converts sunlight into electrical energy using solar cells (solar cells)[1].

### Basic Mechanism
Solar cells are primarily made from semiconductor materials such as silicon, and when sunlight hits them, electrons move due to the photoelectric effect, generating current[1]. Since this current is direct current, it needs to be converted to alternating current using a device called an inverter for use as standard household power[2].

### Main Components
- **Solar panels**: Convert sunlight to electricity[1]
- **Inverter**: Convert DC to AC[2]
- **Power conditioner**: Adjust and control power[2]

### Generation Efficiency
The conversion efficiency of common solar cells today is around 15-20%[1].

## References
[1] Solar Power Basics - New Energy Foundation (https://example.com/solar-basics, April 2023)
[2] Solar Power System Components - Energy Technology Institute (https://example.com/solar-system, March 2023)

---

## Notes
- If search results contain outdated information, please note this
- If there are conflicting pieces of information, present both perspectives
- When using technical terms, include simple explanations whenever possible
"""

STRANDS_KNOWLEDGE_BASE_ID = os.environ.get("STRANDS_KNOWLEDGE_BASE_ID", "XXXXXXXXXX")

@app.entrypoint
async def entrypoint(payload):
  logger.info("Processing started")
  message = payload.get("prompt", "")
  logger.info("prompt=%r", message)
  str_message = str(message)
  logger.info("str_message=%r", str_message)
  model = payload.get("model", {})
  model_id = model.get("modelId","anthropic.claude-3-5-haiku-20241022-v1:0")
  model = BedrockModel(model_id=model_id, params={"max_tokens": 4096, "temperature": 0.7}, region="us-west-2")
  agent = Agent(model=model, system_prompt=SYSTEM_PROMPT, tools=[retrieve_module])

  # Retrieving from Knowledge Base (only if environment variable is valid)
  augmented_prompt = str_message
  kb_id = STRANDS_KNOWLEDGE_BASE_ID
  logger.info("kb_id=%s", kb_id)
  try:
      if kb_id and kb_id != "your_kb_id":
          kb_results = agent.tool.retrieve(
              text=str_message,
              knowledgeBaseId=kb_id,
              numberOfResults=9,
              score=0.4,
              region="us-west-2",
          )
          logger.info("kb_results=%s", kb_results)
          kb_block = str(kb_results)
          logger.info("kb_block=%s", kb_block)
          if kb_block:
              augmented_prompt = (
                  "The following are search results from the Knowledge Base. Please always cite using numbers in your response.\n"
                  "Search Results:\n" + kb_block + "\n\n"
                  "User Question:\n" + str_message
              )
  except Exception as e:
      # Fall back to normal response if KB retrieval fails
      logger.exception("Failed to retrieve from Knowledge Base")
      augmented_prompt = str_message

  stream_messages = agent.stream_async(augmented_prompt)
  async for message in stream_messages:
      if "event" in message:
          yield message

if __name__ == "__main__":
  app.run()

import os
import json
from typing import Optional
import boto3
from strands import Agent, tool
from strands.models import BedrockModel
from bedrock_agentcore.runtime import BedrockAgentCoreApp

app = BedrockAgentCoreApp()

SYSTEM_PROMPT = """
# RAG System Prompt

## Basic Policy
You are an assistant that provides accurate and useful answers based on search results. Please follow these guidelines for your responses.

## Response Rules

### 1. Utilizing Search Results
- Use only the provided search results as information sources
- For information not included in the search results, do not make assumptions or supplement with general knowledge; clearly state "Not included in the search results"
- If there are multiple search results, integrate them to create a comprehensive answer

### 2. Citation Method
- Always include citation numbers for information in your answer (e.g., [1], [2])
- Include a "## References" section after your answer in the following format:
  ```
  [1] Title - Source (URL, date, etc.)
  [2] Title - Source (URL, date, etc.)
  ```
- If search results are provided as JSON, extract the uri from each item and use it as the URL in the references
- For references, clearly indicate "filename and S3 URL". Use the end of the URI as the filename (e.g., s3://bucket/path/file.pdf → file.pdf), and use location.s3Location.uri (or metadata.x-amz-bedrock-kb-source-uri if not available). If there's a page number, add (p.<number>).
  Example:
  ```
  [1] s3://strands-sample-xxx/Amazon Bedrock AgentCoreを使ってみよう！ 〜各種機能のポイントを解説〜 (5).pdf (p.78)
  [2] s3://strands-sample-xxx/Amazon Bedrock AgentCoreを使ってみよう！ 〜各種機能のポイントを解説〜 (5).pdf (p.4)
  ```

### 3. Response Format
- Answer in clear and concise Japanese
- Organize important points using headings and bullet points
- For longer answers, include a summary at the beginning

### 4. When Information is Insufficient
- If search results are insufficient for the question, clearly state this
- If you can provide a partial answer, clearly specify the scope of what you can answer

### 5. Tool Usage
- This agent can use the `kb_search(query, max_results?)` tool
- Call `kb_search` with appropriate queries before answering and use the results as evidence
- When there are multiple results, prioritize those with higher importance and integrate them, avoiding duplication
- Number citations in the order used and match them with the "References" section at the end (including title/URL/date as available)

## Answer Example

**Question:** Please explain how solar power generation works.

**Answer:**

Solar power generation is a system that converts sunlight directly into electrical energy using solar cells (solar cells)[1].

### Basic Mechanism
Solar cells are mainly made of semiconductor materials such as silicon, and when exposed to sunlight, electrons move due to the photoelectric effect, generating an electric current[1]. Since this current is direct current, it needs to be converted to alternating current using a device called an inverter for use as general household power[2].

### Main Components
- **Solar Panels**: Convert sunlight into electricity[1]
- **Inverter**: Convert DC to AC[2]
- **Power Conditioner**: Adjusts and controls power[2]

### Generation Efficiency
The conversion efficiency of typical solar cells today is about 15-20%[1].

## References
[1] Solar Power Generation Basics - New Energy Foundation (https://example.com/solar-basics, April 2023)
[2] Solar Power System Configuration - Energy Technology Institute (https://example.com/solar-system, March 2023)

---

## Notes
- If the search results contain outdated information, please mention this
- If there is contradictory information, present both perspectives
- When using technical terms, provide easy-to-understand explanations whenever possible
"""

STRANDS_KNOWLEDGE_BASE_ID = os.environ.get("STRANDS_KNOWLEDGE_BASE_ID", "XXXXXXXXXX")

# =========================
# Bedrock KB Tool Setup
# =========================
BEDROCK_CLIENT = None
KB_REGION = os.environ.get("AWS_REGION", "us-west-2")
KB_MAX_RESULTS = int(os.environ.get("BEDROCK_KB_MAX_RESULTS", "5"))

def _ensure_bedrock_client(region_name: Optional[str] = None):
    global BEDROCK_CLIENT
    if BEDROCK_CLIENT is None:
        region = region_name or KB_REGION
        BEDROCK_CLIENT = boto3.client("bedrock-agent-runtime", region_name=region)
    return BEDROCK_CLIENT

@tool
def kb_search(query: str, max_results: Optional[int] = None) -> str:
    """
    Search (Retrieve) related documents from the knowledge base.

    Args:
        query: Search query
        max_results: Maximum number of items to retrieve (defaults to environment variable if omitted)

    Returns:
        JSON string (including success, results, etc.)
    """
    try:
        client = _ensure_bedrock_client()
        kb_id = STRANDS_KNOWLEDGE_BASE_ID
        if not kb_id:
            return json.dumps({
                "success": False,
                "error": "Environment variable STRANDS_KNOWLEDGE_BASE_ID is not set",
                "results": [],
            }, ensure_ascii=False)

        num = max_results or KB_MAX_RESULTS
        response = client.retrieve(
            knowledgeBaseId=kb_id,
            retrievalQuery={"text": query},
            retrievalConfiguration={
                "vectorSearchConfiguration": {
                    "numberOfResults": num,
                    "overrideSearchType": "SEMANTIC",
                }
            },
        )

        items = []
        for r in response.get("retrievalResults", []):
            content = r.get("content", {})
            location = r.get("location", {})
            metadata = r.get("metadata", {})
            s3loc = location.get("s3Location", {})
            page = metadata.get("x-amz-bedrock-kb-document-page-number")
            if isinstance(page, float) and page.is_integer():
                page = int(page)
            items.append({
                "content": content.get("text", ""),
                "type": content.get("type"),
                "score": r.get("score", 0),
                "uri": s3loc.get("uri") or metadata.get("x-amz-bedrock-kb-source-uri"),
                "page": page,
                "chunkId": metadata.get("x-amz-bedrock-kb-chunk-id"),
                "dataSourceId": metadata.get("x-amz-bedrock-kb-data-source-id"),
            })
        return json.dumps({
            "success": True,
            "query": query,
            "results_count": len(items),
            "results": items,
        }, ensure_ascii=False)
    except Exception as e:
        return json.dumps({
            "success": False,
            "query": query,
            "error": str(e),
            "results": [],
        }, ensure_ascii=False)

@app.entrypoint
async def entrypoint(payload):
    message = payload.get("prompt", "")
    model = payload.get("model", {})
    model_id = model.get("modelId", "anthropic.claude-3-5-haiku-20241022-v1:0")
    model = BedrockModel(
        model_id=model_id,
        params={"max_tokens": 4096, "temperature": 0.7},
        region=KB_REGION,
    )
    agent = Agent(
        model=model,
        system_prompt=SYSTEM_PROMPT,
        tools=[kb_search],
    )

    stream_messages = agent.stream_async(message)
    async for message in stream_messages:
        if "event" in message:
            yield message

if __name__ == "__main__":
    app.run()

I'll look up information about AgentCore for you.
Let me explain Amazon Bedrock AgentCore.

## Overview

Amazon Bedrock AgentCore is an optimal managed service for deploying and operating AI agents[1]. It was released in public preview on July 16, 2024, and includes all the necessary functions for AI agent development[2].

### Key Features
- **AI Agent Hosting**: Supports various agent frameworks including Strands Agents, LangGraph, and more[3]
- **Convenient Managed Service**: Provides useful managed services for authentication, tool integration, and other functions when using AI agents[3]

## Main Functions

Amazon Bedrock AgentCore has the following six main functions[2]:

### 1. **Runtime**
- Hosting functionality[2]
- Foundational capability for deploying AI agents

### 2. **Identity**
- Access control and authentication for AI agents[2]

### 3. **Gateway**
- Tool creation functionality for external processing[2]
- Enables integration with external services

### 4. **Memory**
- Functionality to store conversation history and state of AI agents[2]

### 5. **Built-in Tools**
AgentCore includes the following convenient built-in tools[2]:
- **Code Interpreter**: Code execution environment
- **Browser**: Browser execution environment

### 6. **Observability**
- Functionality to monitor the operational status of AI agents[1]

## What is an AI Agent?

An AI agent is an AI system that can make judgments and act autonomously. It understands user instructions, plans and uses necessary tools on its own to execute tasks[3].

## Summary

Amazon Bedrock AgentCore is a managed service that enables efficient construction of AI agents using Runtime, Identity, Memory, Gateway, Built-in Tools, and Observability[1]. It supports various agent frameworks and provides integrated functionalities necessary for AI agent development.

## References
[1] Amazon Bedrock AgentCoreを使ってみよう！ 〜各種機能のポイントを解説〜 (5).pdf (p.78) - s3://strands-sample-xxxxx/Amazon Bedrock AgentCoreを使ってみよう！ 〜各種機能のポイントを解説〜 (5).pdf
[2] Amazon Bedrock AgentCoreを使ってみよう！ 〜各種機能のポイントを解説〜 (5).pdf (p.12) - s3://strands-sample-xxxxx/Amazon Bedrock AgentCoreを使ってみよう！ 〜各種機能のポイントを解説〜 (5).pdf  
[3] Amazon Bedrock AgentCoreを使ってみよう！ 〜各種機能のポイントを解説〜 (5).pdf (p.10) - s3://strands-sample-xxxxx/Amazon Bedrock AgentCoreを使ってみよう！ 〜各種機能のポイントを解説〜 (5).pdf