Model-based evaluation
This is a Jupyter notebook

Model-based Evals in Langfuse

Model-based evaluations are a powerful tool to automate the evaluation of production completions in Langfuse.

Currently, model-based evals can be applied to production data in Langfuse via the Python SDK. This gives you full flexibility to run various eval libraries on your production data and discover which work well for your use case.

Coming soon: support for running model-based evals directly from the Langfuse UI/Server.

Example using Langchain Evals

This cookbook shows how model-based evaluations can be used to automate the evaluation of production completions in Langfuse. This example uses Langchain but any other eval library can be used as well. Which library is the best to use depends heavily on the use case.

This cookbook follows three steps:

  1. Fetch production generations stored in Langfuse
  2. Evaluate these generations using Langchain
  3. Ingest results back into Langfuse as scores

Not using Langfuse yet? Get started by capturing LLM events.


First you need to install Langfuse and Langchain via pip and then set the environment variables.

%pip install langfuse langchain openai cohere tiktoken --upgrade
import os
# get keys for your project from
os.environ["LANGFUSE_PUBLIC_KEY"] = ""
os.environ["LANGFUSE_SECRET_KEY"] = ""
# your openai key
os.environ["OPENAI_API_KEY"] = ""
# Your host, defaults to
# For US data region, set to ""
# os.environ["LANGFUSE_HOST"] = "http://localhost:3000"
os.environ['EVAL_MODEL'] = "text-davinci-003"
# Langchain Eval types
    "hallucination": True,
    "conciseness": True,
    "relevance": True,
    "coherence": True,
    "harmfulness": True,
    "maliciousness": True,
    "helpfulness": True,
    "controversiality": True,
    "misogyny": True,
    "criminality": True,
    "insensitivity": True

Initialize the Langfuse Python SDK, more information here (opens in a new tab).

from langfuse import Langfuse
langfuse = Langfuse()

Fetching data

Load all generations from Langfuse filtered by name, in this case OpenAI. Names are used in Langfuse to identify different types of generations within an application. Change it to the name you want to evaluate.

Checkout docs (opens in a new tab) on how to set the name when ingesting an LLM Generation.

def fetch_all_pages(name=None, user_id = None, limit=50):
    page = 1
    all_data = []
    while True:
        response = langfuse.get_generations(name=name, limit=limit, user_id=user_id, page=page)
        if not
        page += 1
    return all_data
generations = fetch_all_pages(user_id='user:abc')

Set up evaluation functions

In this section, we define functions to set up the Langchain eval based on the entries in EVAL_TYPES. Hallucinations require their own function. More on the Langchain evals can be found here (opens in a new tab).

from langchain.evaluation import load_evaluator, EvaluatorType
from langchain import PromptTemplate, OpenAI, LLMChain
from langchain.evaluation.criteria import LabeledCriteriaEvalChain
def get_evaluator_for_key(key: str):
  llm = OpenAI(temperature=0, model=os.environ.get('EVAL_MODEL'))
  return load_evaluator("criteria", criteria=key, llm=llm)
def get_hallucination_eval():
  criteria = {
    "hallucination": (
      "Does this submission contain information"
      " not present in the input or reference?"
  llm = OpenAI(temperature=0, model=os.environ.get('EVAL_MODEL'))
  return LabeledCriteriaEvalChain.from_llm(

Execute evaluation

Below, we execute the evaluation for each Generation loaded above. Each score is ingested into Langfuse via langfuse.score() (opens in a new tab).

from langfuse.model import InitialScore
def execute_eval_and_score():
  for generation in generations:
    criteria = [key for key, value in EVAL_TYPES.items() if value and key != "hallucination"]
    for criterion in criteria:
      eval_result = get_evaluator_for_key(criterion).evaluate_strings(
      langfuse.score(InitialScore(name=criterion, traceId=generation.trace_id,, value=eval_result["score"], comment=eval_result['reasoning']))
# hallucination
def eval_hallucination():
  chain = get_hallucination_eval()
  for generation in generations:
    eval_result = chain.evaluate_strings(
    if eval_result is not None and eval_result["score"] is not None and eval_result["reasoning"] is not None:
      langfuse.score(InitialScore(name='hallucination', traceId=generation.trace_id,, value=eval_result["score"], comment=eval_result['reasoning']))
if EVAL_TYPES.get("hallucination") == True:
# SDK is async, make sure to await all requests

See Scores in Langfuse

In the Langfuse UI, you can filter Traces by Scores and look into the details for each. Check out Langfuse Analytics to understand the impact of new prompt versions or application releases on these scores.

Image of Trace Example trace with conciseness score

Get in touch

Looking for a specific way to score your production data in Langfuse? Join the Discord (opens in a new tab) and discuss your use case!

Was this page useful?

Questions? We're here to help