OpenInference LiteLLM Instrumentation

LiteLLM allows developers to call all LLM APIs using the openAI format. LiteLLM Proxy is a proxy server to call 100+ LLMs in OpenAI format. Both are supported by this auto-instrumentation.

This package implements OpenInference tracing for the following LiteLLM functions:

• completion()
• acompletion()
• completion_with_retries()
• embedding()
• aembedding()
• image_generation()
• aimage_generation()
• anthropic.messages.create()
• anthropic.messages.acreate()

These traces are fully OpenTelemetry compatible and can be sent to an OpenTelemetry collector for viewing, such as Arize Phoenix or Arize AX.

Installation

pip install openinference-instrumentation-litellm

Quickstart

In a notebook environment (jupyter, colab, etc.) install openinference-instrumentation-litellm if you haven't already as well as arize-phoenix and litellm.

pip install openinference-instrumentation-litellm arize-phoenix litellm

First, import dependencies required to autoinstrument liteLLM and set up phoenix as an collector for OpenInference traces.

import litellm
import phoenix as px

from openinference.instrumentation.litellm import LiteLLMInstrumentor

from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import SimpleSpanProcessor

Next, we'll start a phoenix server and set it as a collector.

session = px.launch_app()
endpoint = "http://127.0.0.1:6006/v1/traces"
tracer_provider = TracerProvider()
tracer_provider.add_span_processor(SimpleSpanProcessor(OTLPSpanExporter(endpoint)))

Set up any API keys needed in you API calls. For example:

import os
os.environ["OPENAI_API_KEY"] = "PASTE_YOUR_API_KEY_HERE"

Instrumenting LiteLLM is simple:

LiteLLMInstrumentor().instrument(tracer_provider=tracer_provider)

Now, all calls to LiteLLM functions are instrumented and can be viewed in the phoenix UI.

completion_response = litellm.completion(model="gpt-3.5-turbo", 
                   messages=[{"content": "What's the capital of China?", "role": "user"}])
print(completion_response)

acompletion_response = await litellm.acompletion(
            model="gpt-3.5-turbo",
            messages=[{ "content": "Hello, I want to bake a cake","role": "user"},
                      { "content": "Hello, I can pull up some recipes for cakes.","role": "assistant"},
                      { "content": "No actually I want to make a pie","role": "user"},],
            temperature=0.7,
            max_tokens=20
        )
print(acompletion_response)

embedding_response = litellm.embedding(model='text-embedding-ada-002', input=["good morning!"])
print(embedding_response)

image_gen_response = litellm.image_generation(model='dall-e-2', prompt="cute baby otter")
print(image_gen_response)

You can also uninstrument the functions as follows

LiteLLMInstrumentor().uninstrument(tracer_provider=tracer_provider)

Now any liteLLM function calls you make will not send traces to Phoenix until instrumented again

More Info

• Details on how to setup a LiteLLM Proxy
• More info on OpenInference and Phoenix
• More info on OpenInference and Arize AX
• How to customize spans to track sessions, metadata, etc.
• How to account for private information and span payload customization