Managing AI conversations

Introduction to Amazon Bedrock

Nikhil Rangarajan

Data Scientist

Conversation state management in Bedrock

Managing conversations: essential for natural, coherent AI interactions
Conversation context: includes history and metadata
Claude models can account for context

An illustration showing circles with one element of the conversation each: the message, the history, and the metadata.

Implementing conversation management

ConversationManager class handles core conversation functionality
Message history stores both user and assistant messages with roles
Structured approach ensures scalability and maintainability

class ConversationManager:

  def __init__(self):
    self.bedrock = 
        boto3.client('bedrock-runtime')

    self.conversation_history = []
    self.max_tokens = 8000


  def add_message(self, role, content):
      self.conversation_history
      .append({"role": role,

        "content": content

  })

Using ConversationManager in practice

Add new messages to the conversation log
Ensure the user's input is logged
Help the model maintain context

conversation_manager = ConversationManager()
conversation_manager.add_message("user", "What's the weather like?")

Using ConversationManager in practice

Format as message
Limit the amount of text sent
Send formatted conversation to the model

# Send the latest context to the model
conversation_history = conversation_manager.conversation_history
messages = "\n\n".join(
    f"{msg['role']}: {msg['content']}" 
    for msg in conversation_history[-2:]) # Only most recent exchanges


# Send the full message to the model
response = conversation_manager.bedrock.invoke_model(
    modelId="amazon.nova-micro-v1:0", body=messages)

Let's practice!

Introduction to Amazon Bedrock