Langfuse introduction

Written by: Igor Korotach

What is Langfuse?

Langfuse is an open-source observability and analytics platform for LLM applications

Key capabilities:

 

  • Logging LLM calls (requests, responses, metadata).
  • Python/JS SDK.
  • Tracing user sessions and workflows.
  • Git-like prompt storing.
  • Evaluations (automated + human).
  • Analytics dashboards.
  • Experiments

Why/When is it helpful?

Common LLM challenges

  • Hard to debug LLM behavior.

  • No visibility into prompt versions or outputs.

  • Difficult to evaluate model quality over time.

  • No single interface for LLM observability

  • Difficult to track costs for individual operations

How Langfuse addresses that

  • Faster debugging.

  • Centralized prompt management.

  • Performance insights & monitoring.

  • Enables A/B testing and improvement loops.

  • Is model agnostic and collects recent pricing policies for different models.

  • Enables custom pipelines and integrations.

How it works?

Key Concepts

  • Trace = end-to-end user request (e.g., a conversation, collection of generations).

  • Span = one step inside the trace (e.g., LLM call, database fetch).

  • Prompt versioning = store, reuse, and compare prompts over time.

  • Evaluations = automated scoring or human feedback on outputs.

General architecture

Usage example

from langchain_openai import ChatOpenAI
from langchain.prompts import ChatPromptTemplate
from langfuse.decorators import observe
from langfuse import Langfuse

langfuse = Langfuse(
    public_key="LF_PUBLIC_KEY",
    secret_key="LF_SECRET_KEY",
    host="https://cloud.langfuse.com"  # or self-hosted URL
)

prompt = langfuse.get_prompt(
    name="my-task-prompt",
    label="dev"
)

template = ChatPromptTemplate.from_template(prompt.prompt)
llm = ChatOpenAI(model="gpt-4o-mini")

@observe(as_type="generation", name="explain_with_prompt")
def explain_concept(concept: str) -> str:
    """Explains a concept using a Langfuse-managed prompt and LangChain LLM."""
    chain = template | llm
    response = chain.invoke({"input": concept})
    return response.content

result = explain_concept("Langfuse observability")
print(result)

UI Example

Thanks for your attention.

Questions?

  • Presentation link: https://slides.com/emulebest/langfuse-intro

Langfuse Introduction

By Igor Korotach

Langfuse Introduction

  • 2

More from Igor Korotach