Observability for Watsonx with Datadog

Step 1

Install the Traceloop SDK and initialize it. It will automatically log all calls to Watsonx, with prompts and completions as separate spans.

from ibm_watson_machine_learning.foundation_models import Model

from traceloop.sdk import Traceloop

Traceloop.init(app_name="watsonx_example")


def get_credentials(api_key):
    return {
        "url": "https://us-south.ml.cloud.ibm.com",
        "apikey": api_key,
    }


iam_api_key = os.environ["IAM_API_KEY"]
project_id = os.environ["PROJECT_ID"]

prompt_input = """Calculate result

Input:
what is the capital of China.

Output:
"""

model_id = "meta-llama/llama-2-70b-chat"
parameters = {
    "decoding_method": "sample",
    "max_new_tokens": 60,
    "min_new_tokens": 10,
    "random_seed": 111,
    "temperature": 0.9,
    "top_k": 50,
    "top_p": 1,
    "repetition_penalty": 2
}

model = Model(
    model_id=model_id,
    params=parameters,
    credentials=get_credentials(iam_api_key),
    project_id=project_id
    )

prompt_input = "What is the the captical of China"
generated_response = model.generate(prompt=prompt_input)

Step 2

Configure your Datadog agent to receive traces from OpenLLMetry

# In your values.yaml
otlp:
  receiver:
    protocols:
      http:
        enabled: true
        
# In the environment of your application
TRACELOOP_BASE_URL=http://datadog-agent-hostname:4318

Docs