Un end-to-end ejemplo que muestra cómo crear e invocar agentes de Amazon Bedrock mediante un AWS SDK

Cree e invoque un agente.

REGION = "us-east-1"
ROLE_POLICY_NAME = "agent_permissions"


class BedrockAgentScenarioWrapper:
    """Runs a scenario that shows how to get started using Amazon Bedrock Agents."""

    def __init__(
            self, bedrock_agent_client, runtime_client, lambda_client, iam_resource, postfix
    ):
        self.iam_resource = iam_resource
        self.lambda_client = lambda_client
        self.bedrock_agent_runtime_client = runtime_client
        self.postfix = postfix

        self.bedrock_wrapper = BedrockAgentWrapper(bedrock_agent_client)

        self.agent = None
        self.agent_alias = None
        self.agent_role = None
        self.prepared_agent_details = None
        self.lambda_role = None
        self.lambda_function = None

    def run_scenario(self):
        print("=" * 88)
        print("Welcome to the Amazon Bedrock Agents demo.")
        print("=" * 88)

        # Query input from user
        print("Let's start with creating an agent:")
        print("-" * 40)
        name, foundation_model = self._request_name_and_model_from_user()
        print("-" * 40)

        # Create an execution role for the agent
        self.agent_role = self._create_agent_role(foundation_model)

        # Create the agent
        self.agent = self._create_agent(name, foundation_model)

        # Prepare a DRAFT version of the agent
        self.prepared_agent_details = self._prepare_agent()

        # Create the agent's Lambda function
        self.lambda_function = self._create_lambda_function()

        # Configure permissions for the agent to invoke the Lambda function
        self._allow_agent_to_invoke_function()
        self._let_function_accept_invocations_from_agent()

        # Create an action group to connect the agent with the Lambda function
        self._create_agent_action_group()

        # If the agent has been modified or any components have been added, prepare the agent again
        components = [self._get_agent()]
        components += self._get_agent_action_groups()
        components += self._get_agent_knowledge_bases()

        latest_update = max(component["updatedAt"] for component in components)
        if latest_update > self.prepared_agent_details["preparedAt"]:
            self.prepared_agent_details = self._prepare_agent()

        # Create an agent alias
        self.agent_alias = self._create_agent_alias()

        # Test the agent
        self._chat_with_agent(self.agent_alias)

        print("=" * 88)
        print("Thanks for running the demo!\n")

        if q.ask("Do you want to delete the created resources? [y/N] ", q.is_yesno):
            self._delete_resources()
            print("=" * 88)
            print(
                "All demo resources have been deleted. Thanks again for running the demo!"
            )
        else:
            self._list_resources()
            print("=" * 88)
            print("Thanks again for running the demo!")

    def _request_name_and_model_from_user(self):
        existing_agent_names = [
            agent["agentName"] for agent in self.bedrock_wrapper.list_agents()
        ]

        while True:
            name = q.ask("Enter an agent name: ", self.is_valid_agent_name)
            if name.lower() not in [n.lower() for n in existing_agent_names]:
                break
            print(
                f"Agent {name} conflicts with an existing agent. Please use a different name."
            )

        models = ["anthropic.claude-instant-v1", "anthropic.claude-v2"]
        model_id = models[
            q.choose("Which foundation model would you like to use? ", models)
        ]

        return name, model_id

    def _create_agent_role(self, model_id):
        role_name = f"AmazonBedrockExecutionRoleForAgents_{self.postfix}"
        model_arn = f"arn:aws:bedrock:{REGION}::foundation-model/{model_id}*"

        print("Creating an an execution role for the agent...")

        try:
            role = self.iam_resource.create_role(
                RoleName=role_name,
                AssumeRolePolicyDocument=json.dumps(
                    {
                        "Version": "2012-10-17",
                        "Statement": [
                            {
                                "Effect": "Allow",
                                "Principal": {"Service": "bedrock.amazonaws.com"},
                                "Action": "sts:AssumeRole",
                            }
                        ],
                    }
                ),
            )

            role.Policy(ROLE_POLICY_NAME).put(
                PolicyDocument=json.dumps(
                    {
                        "Version": "2012-10-17",
                        "Statement": [
                            {
                                "Effect": "Allow",
                                "Action": "bedrock:InvokeModel",
                                "Resource": model_arn,
                            }
                        ],
                    }
                )
            )
        except ClientError as e:
            logger.error(f"Couldn't create role {role_name}. Here's why: {e}")
            raise

        return role

    def _create_agent(self, name, model_id):
        print("Creating the agent...")

        instruction = """
            You are a friendly chat bot. You have access to a function called that returns
            information about the current date and time. When responding with date or time,
            please make sure to add the timezone UTC.
            """
        agent = self.bedrock_wrapper.create_agent(
            agent_name=name,
            foundation_model=model_id,
            instruction=instruction,
            role_arn=self.agent_role.arn,
        )
        self._wait_for_agent_status(agent["agentId"], "NOT_PREPARED")

        return agent

    def _prepare_agent(self):
        print("Preparing the agent...")

        agent_id = self.agent["agentId"]
        prepared_agent_details = self.bedrock_wrapper.prepare_agent(agent_id)
        self._wait_for_agent_status(agent_id, "PREPARED")

        return prepared_agent_details

    def _create_lambda_function(self):
        print("Creating the Lambda function...")

        function_name = f"AmazonBedrockExampleFunction_{self.postfix}"

        self.lambda_role = self._create_lambda_role()

        try:
            deployment_package = self._create_deployment_package(function_name)

            lambda_function = self.lambda_client.create_function(
                FunctionName=function_name,
                Description="Lambda function for Amazon Bedrock example",
                Runtime="python3.11",
                Role=self.lambda_role.arn,
                Handler=f"{function_name}.lambda_handler",
                Code={"ZipFile": deployment_package},
                Publish=True,
            )

            waiter = self.lambda_client.get_waiter("function_active_v2")
            waiter.wait(FunctionName=function_name)

        except ClientError as e:
            logger.error(
                f"Couldn't create Lambda function {function_name}. Here's why: {e}"
            )
            raise

        return lambda_function

    def _create_lambda_role(self):
        print("Creating an execution role for the Lambda function...")

        role_name = f"AmazonBedrockExecutionRoleForLambda_{self.postfix}"

        try:
            role = self.iam_resource.create_role(
                RoleName=role_name,
                AssumeRolePolicyDocument=json.dumps(
                    {
                        "Version": "2012-10-17",
                        "Statement": [
                            {
                                "Effect": "Allow",
                                "Principal": {"Service": "lambda.amazonaws.com"},
                                "Action": "sts:AssumeRole",
                            }
                        ],
                    }
                ),
            )
            role.attach_policy(
                PolicyArn="arn:aws:iam::aws:policy/service-role/AWSLambdaBasicExecutionRole"
            )
            print(f"Created role {role_name}")
        except ClientError as e:
            logger.error(f"Couldn't create role {role_name}. Here's why: {e}")
            raise

        print("Waiting for the execution role to be fully propagated...")
        wait(10)

        return role

    def _allow_agent_to_invoke_function(self):
        policy = self.iam_resource.RolePolicy(
            self.agent_role.role_name, ROLE_POLICY_NAME
        )
        doc = policy.policy_document
        doc["Statement"].append(
            {
                "Effect": "Allow",
                "Action": "lambda:InvokeFunction",
                "Resource": self.lambda_function["FunctionArn"],
            }
        )
        self.agent_role.Policy(ROLE_POLICY_NAME).put(PolicyDocument=json.dumps(doc))

    def _let_function_accept_invocations_from_agent(self):
        try:
            self.lambda_client.add_permission(
                FunctionName=self.lambda_function["FunctionName"],
                SourceArn=self.agent["agentArn"],
                StatementId="BedrockAccess",
                Action="lambda:InvokeFunction",
                Principal="bedrock.amazonaws.com",
            )
        except ClientError as e:
            logger.error(
                f"Couldn't grant Bedrock permission to invoke the Lambda function. Here's why: {e}"
            )
            raise

    def _create_agent_action_group(self):
        print("Creating an action group for the agent...")

        try:
            with open("./scenario_resources/api_schema.yaml") as file:
                self.bedrock_wrapper.create_agent_action_group(
                    name="current_date_and_time",
                    description="Gets the current date and time.",
                    agent_id=self.agent["agentId"],
                    agent_version=self.prepared_agent_details["agentVersion"],
                    function_arn=self.lambda_function["FunctionArn"],
                    api_schema=json.dumps(yaml.safe_load(file)),
                )
        except ClientError as e:
            logger.error(f"Couldn't create agent action group. Here's why: {e}")
            raise

    def _get_agent(self):
        return self.bedrock_wrapper.get_agent(self.agent["agentId"])

    def _get_agent_action_groups(self):
        return self.bedrock_wrapper.list_agent_action_groups(
            self.agent["agentId"], self.prepared_agent_details["agentVersion"]
        )

    def _get_agent_knowledge_bases(self):
        return self.bedrock_wrapper.list_agent_knowledge_bases(
            self.agent["agentId"], self.prepared_agent_details["agentVersion"]
        )

    def _create_agent_alias(self):
        print("Creating an agent alias...")

        agent_alias_name = "test_agent_alias"
        agent_alias = self.bedrock_wrapper.create_agent_alias(
            agent_alias_name, self.agent["agentId"]
        )

        self._wait_for_agent_status(self.agent["agentId"], "PREPARED")

        return agent_alias

    def _wait_for_agent_status(self, agent_id, status):
        while self.bedrock_wrapper.get_agent(agent_id)["agentStatus"] != status:
            wait(2)

    def _chat_with_agent(self, agent_alias):
        print("-" * 88)
        print("The agent is ready to chat.")
        print("Try asking for the date or time. Type 'exit' to quit.")

        # Create a unique session ID for the conversation
        session_id = uuid.uuid4().hex

        while True:
            prompt = q.ask("Prompt: ", q.non_empty)

            if prompt == "exit":
                break

            response = asyncio.run(self._invoke_agent(agent_alias, prompt, session_id))

            print(f"Agent: {response}")

    async def _invoke_agent(self, agent_alias, prompt, session_id):
        response = self.bedrock_agent_runtime_client.invoke_agent(
            agentId=self.agent["agentId"],
            agentAliasId=agent_alias["agentAliasId"],
            sessionId=session_id,
            inputText=prompt,
        )

        completion = ""

        for event in response.get("completion"):
            chunk = event["chunk"]
            completion += chunk["bytes"].decode()

        return completion

    def _delete_resources(self):
        if self.agent:
            agent_id = self.agent["agentId"]

            if self.agent_alias:
                agent_alias_id = self.agent_alias["agentAliasId"]
                print("Deleting agent alias...")
                self.bedrock_wrapper.delete_agent_alias(agent_id, agent_alias_id)

            print("Deleting agent...")
            agent_status = self.bedrock_wrapper.delete_agent(agent_id)["agentStatus"]
            while agent_status == "DELETING":
                wait(5)
                try:
                    agent_status = self.bedrock_wrapper.get_agent(
                        agent_id, log_error=False
                    )["agentStatus"]
                except ClientError as err:
                    if err.response["Error"]["Code"] == "ResourceNotFoundException":
                        agent_status = "DELETED"

        if self.lambda_function:
            name = self.lambda_function["FunctionName"]
            print(f"Deleting function '{name}'...")
            self.lambda_client.delete_function(FunctionName=name)

        if self.agent_role:
            print(f"Deleting role '{self.agent_role.role_name}'...")
            self.agent_role.Policy(ROLE_POLICY_NAME).delete()
            self.agent_role.delete()

        if self.lambda_role:
            print(f"Deleting role '{self.lambda_role.role_name}'...")
            for policy in self.lambda_role.attached_policies.all():
                policy.detach_role(RoleName=self.lambda_role.role_name)
            self.lambda_role.delete()

    def _list_resources(self):
        print("-" * 40)
        print(f"Here is the list of created resources in '{REGION}'.")
        print("Make sure you delete them once you're done to avoid unnecessary costs.")
        if self.agent:
            print(f"Bedrock Agent:   {self.agent['agentName']}")
        if self.lambda_function:
            print(f"Lambda function: {self.lambda_function['FunctionName']}")
        if self.agent_role:
            print(f"IAM role:        {self.agent_role.role_name}")
        if self.lambda_role:
            print(f"IAM role:        {self.lambda_role.role_name}")

    @staticmethod
    def is_valid_agent_name(answer):
        valid_regex = r"^[a-zA-Z0-9_-]{1,100}$"
        return (
            answer
            if answer and len(answer) <= 100 and re.match(valid_regex, answer)
            else None,
            "I need a name for the agent, please. Valid characters are a-z, A-Z, 0-9, _ (underscore) and - (hyphen).",
        )

    @staticmethod
    def _create_deployment_package(function_name):
        buffer = io.BytesIO()
        with zipfile.ZipFile(buffer, "w") as zipped:
            zipped.write(
                "./scenario_resources/lambda_function.py", f"{function_name}.py"
            )
        buffer.seek(0)
        return buffer.read()


if __name__ == "__main__":
    logging.basicConfig(level=logging.INFO, format="%(levelname)s: %(message)s")

    postfix = "".join(
        random.choice(string.ascii_lowercase + "0123456789") for _ in range(8)
    )
    scenario = BedrockAgentScenarioWrapper(
        bedrock_agent_client=boto3.client(
            service_name="bedrock-agent", region_name=REGION
        ),
        runtime_client=boto3.client(
            service_name="bedrock-agent-runtime", region_name=REGION
        ),
        lambda_client=boto3.client(service_name="lambda", region_name=REGION),
        iam_resource=boto3.resource("iam"),
        postfix=postfix,
    )
    try:
        scenario.run_scenario()
    except Exception as e:
        logging.exception(f"Something went wrong with the demo. Here's what: {e}")