Integration Mirascope
This is a Jupyter notebook

Cookbook: Mirascope x Langfuse integration

Mirascope (opens in a new tab) is a Python toolkit for building with LLMs. It allows devs to write Pythonic code while profiting from its abstractions to common LLM use cases and models.

Langfuse (opens in a new tab) is an open source LLM engineering platform. Traces, evals, prompt management and metrics to debug and improve your LLM application.

With the Langfuse <-> Mirascope integration (opens in a new tab), you can log your application to Langfuse by adding the @with_langfuse decorator.

Let's dive right in with some examples:

# Install Mirascope and Langfuse
%pip install mirascope[all] langfuse
import os
# Get keys for your project from the project settings page
os.environ["LANGFUSE_PUBLIC_KEY"] = ""
os.environ["LANGFUSE_SECRET_KEY"] = ""
os.environ["LANGFUSE_HOST"] = "" # 🇪🇺 EU region
# os.environ["LANGFUSE_HOST"] = "" # 🇺🇸 US region
# Your openai key
os.environ["OPENAI_API_KEY"] = ""

Log a first simple call

from mirascope.langfuse import with_langfuse
from mirascope.openai import OpenAICall, OpenAICallParams
class GeographyGenius(OpenAICall):
    prompt_template = "What's the capital of {country}?"
    country: str
    call_params = OpenAICallParams(model="gpt-4o", temperature=1)
genius = GeographyGenius(country="Japan")
response =  # logs to langfuse

Example trace (opens in a new tab)

Trace of simple Mirascope execution in Langfuse

Let's make this more complex

We'll use

to create and trace a fun rap battle and group everything into a single trace.

from import ChatCompletionMessageParam
from mirascope.openai import OpenAICall
from langfuse.decorators import observe
class Rapper(OpenAICall):
    prompt_template = """
    SYSTEM: This is a rap battle. You are {person}. Make sure to defend you {position}. Only drop two lines at a time, make them rhyme.
    MESSAGES: {history}
    history: list[ChatCompletionMessageParam] = []
    person: str
    position: str
zuck = Rapper(person="Mark Zuckerberg", position="Open source will win in VR/AR/Visual Computing", history=[])
timapple = Rapper(person="Tim Cook", position="Apple builds the best headsets as we are integrated in software and hardware", history=[])
# utility function to update the history of both rappers
def add_to_history(new_line: str, rapper: str):
    zuck.history += [
        {"role": "assistant" if rapper == "zuck" else "user", "content": new_line},
    timapple.history += [
        {"role": "assistant" if rapper == "timapple" else "user", "content": new_line},
## use the langfuse @observe decorator to log any Python function and wrap all logs within it into a single trace
def rap_battle(lines: int):
  # Make sure that the battle starts of juicy
  add_to_history("Yo wassup Zuck, I hate OSS", "timapple")
  for i in range(lines):
      zuck_line =
      print(f"(Zuck): {zuck_line.content}")
      add_to_history(zuck_line.content, "zuck")
      timapple_line =
      print(f"(Tim Apple): {timapple_line.content}")
      add_to_history(timapple_line.content, "timapple")
  return [item["content"] for item in timapple.history]

Head over to the Langfuse Traces table in Langfuse Cloud (opens in a new tab) to see the entire chat history, token counts, cost, model, latencies and more

Example trace (opens in a new tab)

Trace of complex Mirascope execution in Langfuse

That's a wrap.

There's a lot more you can do:

Was this page useful?

Questions? We're here to help

Subscribe to updates