Thinking tags are coming across as "text" responses with nova pro on Bedrock

[automaton82]

Package

langchain-aws

Checked other resources

• I added a descriptive title to this issue
• I searched the LangChain documentation with the integrated search
• I used the GitHub search to find a similar issue and didn't find it
• I am sure this is a bug and not a question or request for help

Example Code

# Example:
"""
Minimal reproducible example for AWS Bedrock Nova Pro returning <thinking> tags
when using bind_tools().
"""

from langchain_aws import ChatBedrock
from langchain_core.messages import HumanMessage
from langchain_core.tools import tool

# Configure your AWS credentials
AWS_ACCESS_KEY_ID = "your_access_key"
AWS_SECRET_ACCESS_KEY = "your_secret_key"
AWS_REGION = "us-east-1"  # or your region

# Create a simple tool
@tool
def get_population_data(continent: str) -> str:
    """Get population data for a continent."""
    data = {
        "Asia": "3705.03",
        "Africa": "1200.00",
        "Europe": "750.00"
    }
    return f"Population of {continent}: {data.get(continent, 'Unknown')} million"

# Initialize the model
model = ChatBedrock(
    model_id="us.amazon.nova-pro-v1:0",
    aws_access_key_id=AWS_ACCESS_KEY_ID,
    aws_secret_access_key=AWS_SECRET_ACCESS_KEY,
    region=AWS_REGION,
)

# Bind tools to the model
tools = [get_population_data]
llm_with_tools = model.bind_tools(tools)

# Test 1: Without tools (should work fine)
print("=" * 60)
print("Test 1: WITHOUT tools")
print("=" * 60)
response_no_tools = model.invoke([HumanMessage(content="What is the population of Asia?")])
print(f"Response content: {response_no_tools.content}")
print()

# Test 2: With tools bound (will show <thinking> tags in response)
print("=" * 60)
print("Test 2: WITH tools (reproduces the issue)")
print("=" * 60)

messages = [HumanMessage(content="What is the population of Asia? Format as CSV.")]

# First call - may invoke tool
response1 = llm_with_tools.invoke(messages)
print(f"First response - tool_calls: {response1.tool_calls}")
print(f"First response content: {response1.content}")
print()

# If tool was called, execute it and get final response
if response1.tool_calls:
    from langchain_core.messages import ToolMessage
    
    messages.append(response1)
    
    # Execute the tool
    for tool_call in response1.tool_calls:
        tool_result = get_population_data.invoke(tool_call)
        messages.append(ToolMessage(content=str(tool_result), tool_call_id=tool_call["id"]))
    
    # Get final response
    response2 = llm_with_tools.invoke(messages)
    print(f"Final response content:\n{response2.content}")
    print()
    print("ISSUE: Notice the <thinking> tags in the response above ^")
    print()
    
    # Show what the content looks like
    if isinstance(response2.content, list):
        print(f"Content is a list with {len(response2.content)} items:")
        for i, item in enumerate(response2.content):
            print(f"  Item {i}: {type(item)} = {item}")
    else:
        print(f"Content type: {type(response2.content)}")
        print(f"Content value: {response2.content}")

Error Message and Stack Trace (if applicable)

<thinking> The query has been executed and the result is that Asia has the highest total population with 3705.03 million. To format this in CSV, I will use a simple structure with headers "Continent" and "Population (M)". </thinking>

Here is the data in CSV format:
Continent,Population (M)
Asia,3705.03

Description

The issue is the <thinking> tags are not coming in as type:thinking but instead as text which seems wrong. I could write regex to remove them but I think they should be properly categorized.