Docs
SDKs
Python
Decorators

Decorator-based Python Integration

Integrate Langfuse Tracing into your LLM applications with the Langfuse Python SDK using the @observe() decorator.

The SDK supports both synchronous and asynchronous functions, automatically handling traces, spans, and generations, along with key execution details like inputs, outputs and timings. This setup allows you to concentrate on developing high-quality applications while benefitting from observability insights with minimal code. The decorator is fully interoperable with our main integrations (more on this below): OpenAI, Langchain, LlamaIndex.

See the reference (opens in a new tab) for a comprehensive list of all available parameters and methods.

Want more control over the traces logged to Langfuse? Check out the low-level Python SDK.

Overview

Decorator Integration

Example

Simple example (decorator + openai integration)

main.py
from langfuse.decorators import observe
from langfuse.openai import openai # OpenAI integration
 
@observe()
def story():
    return openai.chat.completions.create(
        model="gpt-3.5-turbo",
        max_tokens=100,
        messages=[
          {"role": "system", "content": "You are a great storyteller."},
          {"role": "user", "content": "Once upon a time in a galaxy far, far away..."}
        ],
    ).choices[0].message.content
 
@observe()
def main():
    return story()
 
main()

Trace in Langfuse (public link (opens in a new tab))

Simple OpenAI decorator trace

Installation & setup

Install the Langfuse Python SDK

PyPI (opens in a new tab)

pip install langfuse

Add Langfuse API keys

If you haven't done so yet, sign up to Langfuse (opens in a new tab) and obtain your API keys from the project settings. Alternatively, you can also run Langfuse locally or self-host.

import os
 
os.environ["LANGFUSE_SECRET_KEY"] = "sk-lf-..."
os.environ["LANGFUSE_PUBLIC_KEY"] = "pk-lf-..."
os.environ["LANGFUSE_HOST"] = "https://cloud.langfuse.com"  # 🇪🇺 EU region
# os.environ["LANGFUSE_HOST"] = "https://us.cloud.langfuse.com" # 🇺🇸 US region

When no API keys are provided, a single warning is logged, and no traces are sent to Langfuse.

Add the Langfuse decorator

Import the @observe() decorator and apply it to the functions you want to trace. By default it captures:

  • nesting via context vars
  • timings/durations
  • function name
  • args and kwargs as input dict
  • returned values as output

The decorator will automatically create a trace for the top-level function and spans for any nested functions. Learn more about the tracing data model here.

from langfuse.decorators import observe
 
@observe()
def fn():
    pass
 
@observe()
def main():
    fn()
 
main()

Done! ✨ Read on to learn how to capture additional information, LLM calls, and more with Langfuse Python decorators.

⚠️

In a short-lived environment like AWS Lambda, make sure to call flush() before the function terminates to avoid losing events. Learn more.

from langfuse.decorators import observe, langfuse_context
 
@observe()
def main():
    print("Hello, from the main function!")
 
main()
 
langfuse_context.flush()

Decorator arguments

See SDK reference (opens in a new tab) for full details.

Log any LLM call

In addition to the native intgerations with LangChain, LlamaIndex, and OpenAI (details below), you can log any LLM call by decorating it with @observe(as_type="generation"). Important: Make sure the as_type="generation" decorated function is called inside another @observe()-decorated function for it to have a top-level trace.

Optionally, you can parse some of the arguments to the LLM call and pass them to langfuse_context.update_current_observation to enrich the trace.

main.py
from langfuse.decorators import observe, langfuse_context
import anthropic
 
anthopic_client = anthropic.Anthropic()
 
# Wrap LLM function with decorator
@observe(as_type="generation")
def anthropic_completion(**kwargs):
  # optional, extract some fields from kwargs
  kwargs_clone = kwargs.copy()
  input = kwargs_clone.pop('messages', None)
  model = kwargs_clone.pop('model', None)
  langfuse_context.update_current_observation(
      input=input,
      model=model,
      metadata=kwargs_clone
  )
 
  response = anthopic_client.messages.create(**kwargs)
 
  # See docs for more details on token counts and usd cost in Langfuse
  # https://langfuse.com/docs/model-usage-and-cost
  langfuse_context.update_current_observation(
      usage={
          "input": response.usage.input_tokens,
          "output": response.usage.output_tokens
      }
  )
 
  # return result
  return response.content[0].text
 
@observe()
def main():
  return anthropic_completion(
      model="claude-3-opus-20240229",
      max_tokens=1024,
      messages=[
          {"role": "user", "content": "Hello, Claude"}
      ]
  )
 
main()

Capturing of input/output

By default, the @observe() decorator captures the input arguments and output results of the function.

You can disable this behavior by setting the capture_input and capture_output parameters to False.

The decorator implementation supports capturing any serializable object as input and output such as strings, numbers, lists, dictionaries, and more. Python generators which are common when streaming LLM responses are supported as return values from decorated functions, but not as input arguments.

from langfuse.decorators import observe
 
@observe(capture_input=False, capture_output=False)
def fn(secret_arg):
    return "super secret output"
 
fn("my secret arg")

You can manually set the input and output of the observation using langfuse_context.update_current_observation (details below).

from langfuse.decorators import langfuse_context, observe
 
@observe(capture_input=False, capture_output=False)
def fn(secret_arg):
    langfuse_context.update_current_observation(
        input="sanitized input", # any serializable object
        output="sanitized output", # any serializable object
    )
    return "super secret output"
 
fn("my secret arg")

This will result in a trace with only sanitized input and output, and no actual function arguments or return values.

Decorator context

Use the langfuse_context object to interact with the decorator context. This object is a thread-local singleton and can be accessed from anywhere within the function context.

Configure the Langfuse client

The decorator manages the Langfuse client for you. If you need to configure the client, you can do so via the langfuse_context.configure method at the top of your application before executing any decorated functions.

from langfuse.decorators import langfuse_context
 
# Configure the Langfuse client
langfuse_context.configure(
    secret_key="sk-lf-...",
    public_key="pk-lf-...",
    httpx_client=custom_httpx_client,
    host=custom_host,
    enabled=True,
)

By setting the enabled parameter to False, you can disable the decorator and prevent any traces from being sent to Langfuse.

See the API Reference (opens in a new tab) for more details on the available parameters.

Add additional attributes to the trace and observations

In addition to the attributes automatically captured by the decorator, you can add others to use the full features of Langfuse.

Please read the reference for more details on available parameters:

Below is an example demonstrating how to enrich traces and observations with custom parameters:

from langfuse.decorators import langfuse_context, observe
 
@observe()
def deeply_nested_fn():
    # Enrich the current observation with a custom name, input, and output
    # All of these parameters override the default values captured by the decorator
    langfuse_context.update_current_observation(
        name="Deeply nested LLM call",
        input="Ping?",
        output="Pong!"
    )
    # Updates the trace, overriding the default trace name `main` (function name)
    langfuse_context.update_current_trace(
        name="Trace name set from deeply_nested_llm_call",
        session_id="1234",
        user_id="5678",
        tags=["tag1", "tag2"],
        public=True
    )
    return "output" # This output will not be captured as we have overridden it
 
@observe()
def nested_fn():
    # Update the current span with a custom name and level
    # Overrides the default span name
    langfuse_context.update_current_observation(
        name="Nested Span",
        level="WARNING"
    )
    deeply_nested_fn()
 
@observe()
def main():
    # This will be the trace as it is the highest level function
    nested_fn()
 
# Execute the main function to generate the enriched trace
main()

Get trace URL

You can get the URL of the current trace using langfuse_context.get_current_trace_url(). Works anywhere within the function context, also in deeply nested functions.

from langfuse.decorators import langfuse_context, observe
 
@observe()
def main():
    print(langfuse_context.get_current_trace_url())
 
main()

Trace/observation IDs

By default, Langfuse assigns random IDs to all logged events.

Get trace and observation IDs

You can access the current trace and observation IDs from the langfuse_context object.

from langfuse.decorators import langfuse_context, observe
 
@observe()
def fn():
    print(langfuse_context.get_current_trace_id())
    print(langfuse_context.get_current_observation_id())
 
fn()

Set custom IDs

If you have your own unique ID (e.g. messageId, traceId, correlationId), you can easily set those as trace or observation IDs for effective lookups in Langfuse. Just pass the langfuse_observation_id keyword argument to the decorated function.

from langfuse.decorators import langfuse_context, observe
 
@observe()
def process_user_request(user_id, request_data, **kwargs):
    # Function logic here
    pass
 
@observe(**kwargs)
def main():
    process_user_request(
        "user_id",
        "request",
        langfuse_observation_id="my-custom-request-id",
    )
 
 
main(langfuse_observation_id="my-custom-request-id")

Interoperability with framework integrations

The decorator is fully interoperable with our main integrations: OpenAI, Langchain, LlamaIndex. Thereby you can easily trace and evaluate functions that use (a combination of) these integrations.

OpenAI

The drop-in OpenAI SDK integration is fully compatible with the @observe() decorator. It automatically adds a generation observation to the trace within the current context.

from langfuse.decorators import observe
from langfuse.openai import openai
 
@observe()
def story():
    return openai.chat.completions.create(
        model="gpt-3.5-turbo",
        max_tokens=100,
        messages=[
          {"role": "system", "content": "You are a great storyteller."},
          {"role": "user", "content": "Once upon a time in a galaxy far, far away..."}
        ],
    ).choices[0].message.content
 
@observe()
def main():
    return story()
 
main()

LangChain

The native LangChain integration is fully compatible with the @observe() decorator. It automatically adds a generation to the trace within the current context.

langfuse_context.get_current_langchain_handler() exposes a callback handler scoped to the current trace context. Pass it to subsequent runs to your LangChain application to get full tracing within the scope of the current trace.

from operator import itemgetter
from langchain_openai import ChatOpenAI
from langchain.prompts import ChatPromptTemplate
from langchain.schema import StrOutputParser
from langfuse.decorators import observe
 
prompt = ChatPromptTemplate.from_template("what is the city {person} is from?")
model = ChatOpenAI()
chain = prompt | model | StrOutputParser()
 
@observe()
def langchain_fn(person: str):
    # Get Langchain Callback Handler scoped to the current trace context
    langfuse_handler = langfuse_context.get_current_langchain_handler()
 
    # Pass handler to invoke of your langchain chain/agent
    chain.invoke({"person": person}, config={"callbacks":[langfuse_handler]})
 
langchain_fn("John Doe")

LlamaIndex

The LlamaIndex integration is fully compatible with the @observe() decorator. It automatically adds a generation to the trace within the current context.

Via Settings.callback_manager you can configure the callback to use for tracing of the subsequent LlamaIndex executions. langfuse_context.get_current_llama_index_handler() exposes a callback handler scoped to the current trace context.

from langfuse.decorators import langfuse_context, observe
from llama_index.core import Document, VectorStoreIndex
from llama_index.core import Settings
from llama_index.core.callbacks import CallbackManager
 
doc1 = Document(text="""
Maxwell "Max" Silverstein, a lauded movie director, screenwriter, and producer, was born on October 25, 1978, in Boston, Massachusetts. A film enthusiast from a young age, his journey began with home movies shot on a Super 8 camera. His passion led him to the University of Southern California (USC), majoring in Film Production. Eventually, he started his career as an assistant director at Paramount Pictures. Silverstein's directorial debut, “Doors Unseen,” a psychological thriller, earned him recognition at the Sundance Film Festival and marked the beginning of a successful directing career.
""")
doc2 = Document(text="""
Throughout his career, Silverstein has been celebrated for his diverse range of filmography and unique narrative technique. He masterfully blends suspense, human emotion, and subtle humor in his storylines. Among his notable works are "Fleeting Echoes," "Halcyon Dusk," and the Academy Award-winning sci-fi epic, "Event Horizon's Brink." His contribution to cinema revolves around examining human nature, the complexity of relationships, and probing reality and perception. Off-camera, he is a dedicated philanthropist living in Los Angeles with his wife and two children.
""")
 
@observe()
def llama_index_fn(question: str):
    # Set callback manager for LlamaIndex, will apply to all LlamaIndex executions in this function
    langfuse_handler = langfuse_context.get_current_llama_index_handler()
    Settings.callback_manager = CallbackManager([langfuse_handler])
 
    # Run application
    index = VectorStoreIndex.from_documents([doc1,doc2])
    response = index.as_query_engine().query(question)
    return response

Adding scores

Scores (opens in a new tab) are used to evaluate single observations or entire traces. They can be created via our annotation workflow in the Langfuse UI or via the SDKs.

ParameterTypeOptionalDescription
namestringnoIdentifier of the score.
valuenumbernoThe value of the score. Can be any number, often standardized to 0..1
commentstringyesAdditional context/explanation of the score.

You can attach a score to the current observation context by calling langfuse_context.score_current_observation. You can also score the entire trace from anywhere inside the nesting hierarchy by calling langfuse_context.score_current_trace:

from langfuse.decorators import langfuse_context, observe
 
# This will create a new span under the trace
@observe()
def nested_span():
    langfuse_context.score_current_observation(
        name="feedback-on-span",
        value=1,
        comment="I like how personalized the response is",
    )
 
    langfuse_context.score_current_trace(
        name="feedback-on-trace",
        value=1,
        comment="I like how personalized the response is",
    )
 
 
# This will create a new trace
@observe()
def main():
    nested_span()
 
main()

Additional configuration

Flush observations

The Langfuse SDK executes network requests in the background on a separate thread for better performance of your application. This can lead to lost events in short lived environments such as AWS Lambda functions when the Python process is terminated before the SDK sent all events to our backend.

To avoid this, ensure that the langfuse_context.flush() method is called before termination. This method is waiting for all tasks to have completed, hence it is blocking.

Debug mode

Enable debug mode to get verbose logs. Set the debug mode via the environment variable LANGFUSE_DEBUG=True.

Sampling

Sampling can be controlled via the LANGFUSE_SAMPLE_RATE environment variable. See the sampling documentation for more details.

Authentication check

Use langfuse_context.auth_check() to verify that your host and API credentials are valid. This operation is blocking and is not recommended for production use.

Limitations

Using ThreadPoolExecutors or ProcessPoolExecutors

The decorator uses Python's contextvars to store the current trace context and to ensure that the observations are correctly associated with the current execution context. However, when using Python's ThreadPoolExecutors and ProcessPoolExecutors and when spawning threads from inside a trace (i.e. the executor is run inside a decorated function) the decorator will not work correctly as the contextvars are not correctly copied to the new threads or processes. There is an existing issue (opens in a new tab) in Python's standard library and a great explanation (opens in a new tab) in the fastapi repo that discusses this limitation.

For example when a @observe-decorated function uses a ThreadPoolExecutor to make concurrent LLM requests the context that holds important info on the nesting hierarchy ("we are inside another trace") is not copied over correctly to the child threads. So the created generations will not be linked to the trace and be 'orphaned'. In the UI, you will see a trace missing those generations.

A workaround is to manually copy over the context to the new threads or processes. This is the recommmended approach when using ThreadPoolExecutors or ProcessPoolExecutors directly:

from concurrent.futures import ThreadPoolExecutor, as_completed
from contextvars import copy_context
 
from langfuse.decorators import langfuse_context, observe
 
@observe()
def execute_task(*args):
    return args
 
task_args = [["a", "b"], ["c", "d"]]
 
@observe()
def execute_groups(task_args):
    with ThreadPoolExecutor(3) as executor:
        futures = []
 
        for task_arg in task_args:
            ctx = copy_context()
            task = lambda p=task_arg: ctx.run(execute_task, *p)
 
            futures.append(executor.submit(task))
 
        for future in as_completed(futures):
            future.result()
 
    return [f.result() for f in futures]
 
execute_groups(task_args)
 
langfuse_context.flush()

The executions inside the ThreadPoolExecutor will now be correctly associated with the trace opened by the execute_groups function.

When spawning threads manually with threading.Thread rather than via ThreadPoolExecutor, contextvars are copied over correctly as no executor is used. The decorator works as intended in this case.

Large input/output data

Large input/output data can lead to performance issues. We recommend disabling capturing input/output for these methods and manually add the relevant information via langfuse_context.update_current_observation.

API reference

See the Python SDK API reference (opens in a new tab) for more details.

Was this page useful?

Questions? We're here to help

Subscribe to updates