OpenTelemetry Cohere Instrumentation

Warnings

• breaking Breaking changes related to OpenTelemetry GenAI Semantic Conventions (gen_ai) have been introduced in versions 0.55.0 and later. Attribute names for LLM-related spans may have changed to align with the new conventions (e.g., `gen_ai.request.model` instead of older variations).
  Fix: Review the official OpenTelemetry 'gen_ai' semantic conventions documentation for updated attribute names. Update your queries, dashboards, and any code relying on specific attribute keys. Consider using `OTEL_SEMCONV_STABILITY_OPT_IN` if available for temporary backward compatibility during migration (check core OTel Python documentation).
• gotcha By default, this instrumentation logs sensitive data like prompts, completions, and embeddings to span attributes, which might contain highly sensitive user data.
  Fix: To disable logging of content, set the environment variable `TRACELOOP_TRACE_CONTENT=false` before running your application.
• gotcha The `CohereInstrumentor().instrument()` call must be executed *before* any `cohere.Client()` instance is created or any Cohere API calls are made, as instrumentation typically relies on monkey-patching existing libraries.
  Fix: Ensure `CohereInstrumentor().instrument()` is called early in your application's startup, preferably before initializing the Cohere client.
• gotcha Simply installing `opentelemetry-instrumentation-cohere` and calling `instrument()` will not make traces visible. You must also configure an OpenTelemetry SDK (e.g., `TracerProvider`, `SpanProcessor`, and an `Exporter`) in your application.
  Fix: Set up a complete OpenTelemetry SDK pipeline, including a `TracerProvider`, at least one `SpanProcessor` (e.g., `SimpleSpanProcessor` or `BatchSpanProcessor`), and an `Exporter` (e.g., `ConsoleSpanExporter`, `OTLPExporter`) to send your telemetry data to a backend or console. Refer to the OpenTelemetry Python SDK documentation for detailed setup.

Install

• pip install opentelemetry-instrumentation-cohere cohere opentelemetry-sdk Install with dependencies

Imports

• CohereInstrumentor

  from opentelemetry.instrumentation.cohere import CohereInstrumentor

  This is the main class to enable Cohere instrumentation.

Quickstart

This quickstart demonstrates how to set up OpenTelemetry tracing for Cohere API calls. It initializes a basic OpenTelemetry SDK with a console exporter, instruments the Cohere library, and then makes a sample chat completion request. Ensure the `COHERE_API_KEY` environment variable is set before running.

import os
import cohere
from opentelemetry import trace
from opentelemetry.sdk.resources import Resource
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import ConsoleSpanExporter, SimpleSpanProcessor
from opentelemetry.instrumentation.cohere import CohereInstrumentor

# 1. Set up OpenTelemetry SDK for console output
resource = Resource.create({"service.name": "my-cohere-app"})
provider = TracerProvider(resource=resource)
processor = SimpleSpanProcessor(ConsoleSpanExporter())
provider.add_span_processor(processor)
trace.set_tracer_provider(provider)

# 2. Instrument Cohere
CohereInstrumentor().instrument()

# 3. Use Cohere client (ensure COHERE_API_KEY environment variable is set)
cohere_api_key = os.environ.get('COHERE_API_KEY', '')

if not cohere_api_key:
    print("Warning: COHERE_API_KEY environment variable not set. Cohere client may fail.")
    # For demonstration, you might want to uncomment and replace 'YOUR_COHERE_API_KEY' for testing
    # cohere_api_key = 'YOUR_COHERE_API_KEY'

if cohere_api_key:
    try:
        client = cohere.Client(cohere_api_key)
        response = client.chat(
            model="command",
            message="Hello, how are you today?"
        )
        print("\nCohere Response:", response.text)
        print("\nTraces should be printed to the console above.")
    except Exception as e:
        print(f"Error calling Cohere: {e}")
else:
    print("Skipping Cohere call because API key is not available.")