🚀 Feature: Add Azure Application Insights as an open telemetry destination about openllmetry

[nirga]

Not yet, hopefully soon :)

[aavetis]

here's the setup i used, and everything "just works" 😄

config.py

from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from azure.monitor.opentelemetry.exporter import AzureMonitorTraceExporter
from opentelemetry.sdk.trace.export import SimpleSpanProcessor

# Set the tracer provider
trace.set_tracer_provider(TracerProvider())

# Configure the tracer provider to export traces to Azure Application Insights
exporter = AzureMonitorTraceExporter(
    connection_string = "$INSTRUMENTATION_KEY_HERE"
)
span_processor = SimpleSpanProcessor(exporter)
trace.get_tracer_provider().add_span_processor(span_processor)

app.py

import config  # Import the configuration file
from traceloop.sdk import Traceloop
from traceloop.sdk.decorators import workflow
from opentelemetry import trace
import openai  # Ensure you have the openai library installed

Traceloop.init(app_name="your_app_name")
tracer = trace.get_tracer(__name__)

@workflow(name="llm_execution")
def execute_llm():
    with tracer.start_as_current_span("llm_span"):
        # Replace with your actual OpenAI API call
        response = openai.ChatCompletion.create(
            model="gpt-3.5-turbo",
            messages=[{"role": "user", "content": "checking if this is logged"}],
            max_tokens=60
        )
        print (response['choices'][0]['message']['content'])
        return response['choices'][0]['message']['content']

if __name__ == "__main__":
    execute_llm()

Resulting in:

---

one thing - it seems that all telemetry goes through Traceloop (the saas) before landing in the final destination. is there a way to turn that off? while personally i think that's very cool how you can setup demo dashboards for people to get an idea of how great the platform is, it's a concern for any real production scenario. definitely want to keep your LLM logs and traces isolated.

[nirga]

Sounds good @aavetis! Indeed it makes sense.
Let me know once you verified that it works and update the docs in this repo (under /docs) and the README. We'll then announce that support with proper credit 😄

[nirga]

Can you try passing your exporter to the sdk's init function? It just that we have an exporter we define internally in the sdk so that's what you're seeing - traces are sent to both destinations in parallel.

[aavetis]

Confirmed, when I explicitly pass in the config, it ONLY shows up in Azure and not in Traceloop dashboard. However, it appears the open dashboard is created anyway, just stays empty if the exporter is set.

# this appears to ensure traces are only sent to your desired exporter
exporter = AzureMonitorTraceExporter(connection_string="InstrumentationKey=.......")
Traceloop.init(app_name="your_app_name", exporter=exporter)

As you can see, the dashboard is still created (just a note).

[nirga]

Huh! That's a bug. Fixed in #96.
@aavetis can you add the details of how you set this up in the docs?

[aavetis]

will work on getting in the docs.

on another note, are Metrics / MeterProvider expected to work? i've noticed traces, spans, and dependencies are all tracked seamlessly.

but i haven't been able to see custom metrics (counters) getting passed down, not sure if i'm making a mistake or if sdk isn't expected to support it.
https://opentelemetry.io/docs/specs/otel/metrics/api/#meter-operations

[drewby]

SimpleSpanProcessor

It may be better to document use of APPLICATIONINSIGHTS_CONNECTION_STRING env variable as a good practice. I believe you can then even omit it from the exporter constructor to simplify the code.

Traceloop.init(app_name="your_app_name", exporter=AzureMonitorTraceExporter())

[nirga]

FYI @aavetis docs have been moved to a separate repo - https://github.com/traceloop/docs
(To support the fact that we now have a JS SDK as well)

[nirga]

@aavetis any update on this?

[aavetis]

traceloop/docs#2

[aavetis]

Closing as traceloop/docs#2 is merged and deployed