Chat with Bedrock in the Amazon Connect contact control panel 🎧 🤖

Services used

Let's build

Prerequisites

1. Create a dynamoDB table with the next partition key: "contact_id" (String) and the rest default settings (see Create a table example).
2. Check the Amazon Connect instance ID and contact flow ID (see Find the flow ID). This will be the Connect contact flow that will receive the Bedrock conversation. Check also that your Connect agents have access to this contact flow and its queue (see How routing works).
   • If you don't have yet an Amazon Connect instance check the Get started with Amazon Connect guide. In just some minutes you can start your own contact center instance in the cloud!.
3. Create a standard SNS topic as explained in the real-time chat message streaming guide step1. (FIFO topics don't work with this!!).
4. Create an Amazon Bedrock agent (see Create an agent in Amazon Bedrock guide).
5. Create a SQS queue as the destination of the SNS topic. (see Fanout to Amazon SQS queues):
   • Set the Visibility timeout to 60 seconds, you can leave the rest default parameters.
   • In the SNS console select the SQS subscription and add the next message attribute policy:

Orchestration lambda

1. Increase timeout value to over 1min. Most of generative AI models will take several seconds to elaborate the answer (see Configure Lambda function timeout).
2. Add lambda role next permissions policies: Bedrock access, Amazon connect access, Dynamodb access and Amazon SQS access (see Lambda execution role and AWS managed policies).
3. Add the previously created SQS queue as the lambda trigger with the next configuration: batch size = 1.
4. Create the next environment variables with your own parameters. You can get the agent_id and alias_id from bedrock agent console, the contact_flow_id and instance_id from Amazon Connect console, the dynamo_table name from DynamoDB and the sns_endpoint_arn from the SNS topic console:

• Start a new Connect conversation. This step is only executed if the message doesn't comes from the SQS trigger as this means that is a new conversation (See if-else logic in the lambda handler). In the code this is done with the function: start_agent_conversation.
• Saving and reading the connection_token (for the Amazon Connect conversation) and the session_id (For the Bedrock conversation). We have to match this two parameters to be able to have the same conversation in the Connect panel and in the Bedrock session. We use a DynamoDB table to store these values ​​persistently. In the code this is done with the functions: save_conversation_id and get_session_and_token.
• Send the Connect response to the Bedrock model. In the code this is done with the function: start_agent_conversation.
• Send the Bedrock response to Connect. In the code this is done with the send_message Api call.

Trigger the lambda and receive the call in Connect

Conclusions

Code references

• Amazon Bedrock Workshop
• How to build a Bedrock agent
• Enable real-time chat message streaming in Amazon Connect

• Bedrock customer use cases
• Build your first generative AI application with Amazon Bedrock video
• Developing Generative AI Applications on AWS course