Langfuse is an open source LLM engineering platform which helps teams collaboratively debug, analyze and iterate on their LLM applications. Its core features are observability (tracing), prompt management (versioning), evaluations (scores) and datasets (testing).

Langfuse allows users to score individual executions or traces. Users can customize the scores and scales they use. As such, UpTrain’s evaluations can easily be integrated into Langfuse.

Scores can be used in a variety of ways in Langfuse:

  1. Data: Attach scores to executions and traces and view them in the Langfuse UI
  2. Filter: Group executions or traces by scores to e.g. filter for traces with a low-quality score
  3. Fine Tuning: Filter and export by scores as .csv or .JSONL for fine-tuning
  4. Analytics: Detailed score reporting and dashboards with drill downs into use cases and user segments

In this guide, we will walk you through using Langfuse to create traces and evaluate them using UpTrain.

​
How to integrate?

​
Setup

Enter your Langfuse API keys and OpenAI API key

You need to sign up to Langfuse and fetch your Langfuse API keys in your project’s settings. You also need an OpenAI API key

%pip install langfuse datasets uptrain litellm openai --upgrade

import os

# get keys for your project from https://cloud.langfuse.com
os.environ["LANGFUSE_PUBLIC_KEY"] = ""
os.environ["LANGFUSE_SECRET_KEY"] = ""

# your openai key
os.environ["OPENAI_API_KEY"] = ""

# Your host, defaults to https://cloud.langfuse.com
# For US data region, set to "https://us.cloud.langfuse.com"
# os.environ["LANGFUSE_HOST"] = "http://localhost:3000"

Let’s create a sample data

data = [
    {
        "question": "What are the symptoms of a heart attack?",
        "context": "A heart attack, or myocardial infarction, occurs when the blood supply to the heart muscle is blocked. Chest pain is a good symptom of heart attack, though there are many others.",
        "response": "Symptoms of a heart attack may include chest pain or discomfort, shortness of breath, nausea, lightheadedness, and pain or discomfort in one or both arms, the jaw, neck, or back."
    },
    {
        "question": "Can stress cause physical health problems?",
        "context": "Stress is the body's response to challenges or threats. Yes, chronic stress can contribute to various physical health problems, including cardiovascular issues.",
        "response": "Yes, chronic stress can contribute to various physical health problems, including cardiovascular issues, and a weakened immune system."
    },
    {
        'question': "What are the symptoms of a heart attack?",
        'context': "A heart attack, or myocardial infarction, occurs when the blood supply to the heart muscle is blocked. Symptoms of a heart attack may include chest pain or discomfort, shortness of breath and nausea.",
        'response': "Heart attack symptoms are usually just indigestion and can be relieved with antacids."
    },
    {
        'question': "Can stress cause physical health problems?",
        'context': "Stress is the body's response to challenges or threats. Yes, chronic stress can contribute to various physical health problems, including cardiovascular issues.",
        'response': "Stress is not real, it is just imaginary!"
    }
]

Run Evaluations using UpTrain Open-Source Software (OSS)

We have used the following 3 metrics from UpTrain’s library:

  1. Context Relevance: Evaluates how relevant the retrieved context is to the question specified.

  2. Factual Accuracy: Evaluates whether the response generated is factually correct and grounded by the provided context.

  3. Response Completeness: Evaluates whether the response has answered all the aspects of the question specified

You can look at the complete list of UpTrain’s supported metrics here

from uptrain import EvalLLM, Evals
import json
import pandas as pd

eval_llm = EvalLLM(openai_api_key=os.environ["OPENAI_API_KEY"])

res = eval_llm.evaluate(
    data = data,
    checks = [Evals.CONTEXT_RELEVANCE, Evals.FACTUAL_ACCURACY, Evals.RESPONSE_COMPLETENESS]
)

​
Using Langfuse

You can use Langfuse in 2 ways:

  1. Score each Trace: This means you will run the evaluations for each trace item. This gives you much better idea since of how each call to your UpTrain pipelines is performing but can be expensive
  2. Score as Batch: In this method we will take a random sample of traces on a periodic basis and score them. This brings down cost and gives you a rough estimate the performance of your app but can miss out on important samples.

​
Method 1: Score with Trace

Now lets initialize a Langfuse client SDK to instrument you app

from langfuse import Langfuse
 
langfuse = Langfuse()

langfuse.auth_check()

Let’s create a trace for the dataset

# start a new trace when you get a question
question = data[0]['question']
trace = langfuse.trace(name = "uptrain trace")

# retrieve the relevant chunks
# chunks = get_similar_chunks(question)
context = data[0]['context']
# pass it as span
trace.span(
    name = "retrieval", input={'question': question}, output={'context': context}
)

# use llm to generate a answer with the chunks
# answer = get_response_from_llm(question, chunks)
response = data[0]['response']
trace.span(
    name = "generation", input={'question': question, 'context': context}, output={'response': response}
)

Let’s add the scores to the trace in Langfuse

trace.score(name='context_relevance', value=res[0]['score_context_relevance'])
trace.score(name='factual_accuracy', value=res[0]['score_factual_accuracy'])
trace.score(name='response_completeness', value=res[0]['score_response_completeness'])

​
Method 2: Scoring as batch

Let’s create trace with our original dataset

for interaction in data:
    trace = langfuse.trace(name = "uptrain batch")
    trace.span(
        name = "retrieval",
        input={'question': interaction['question']},
        output={'context': interaction['context']}
    )
    trace.span(
        name = "generation",
        input={'question': interaction['question'], 'context': interaction['context']},
        output={'response': interaction['response']}
    )
 
# await that Langfuse SDK has processed all events before trying to retrieve it in the next step
langfuse.flush()

Retrieve the uploaded dataset

def get_traces(name=None, limit=10000, user_id=None):
    all_data = []
    page = 1
 
    while True:
        response = langfuse.client.trace.list(
            name=name, page=page, user_id=user_id, order_by=None
        )
        if not response.data:
            break
        page += 1
        all_data.extend(response.data)
        if len(all_data) > limit:
            break
 
    return all_data[:limit]

Now lets make a batch and score it using UpTrain.

from random import sample
 
NUM_TRACES_TO_SAMPLE = 4
traces = get_traces(name="uptrain batch")
traces_sample = sample(traces, NUM_TRACES_TO_SAMPLE)

# score on a sample
from random import sample
 
evaluation_batch = {
    "question": [],
    "context": [],
    "response": [],
    "trace_id": [],
}
 
for t in traces_sample:
    observations = [langfuse.client.observations.get(o) for o in t.observations]
    for o in observations:
        if o.name == 'retrieval':
            question = o.input['question']
            context = o.output['context']
        if o.name=='generation':
            answer = o.output['response']
    evaluation_batch['question'].append(question)
    evaluation_batch['context'].append(context)
    evaluation_batch['response'].append(response)
    evaluation_batch['trace_id'].append(t.id)

data = [dict(zip(evaluation_batch,t)) for t in zip(*evaluation_batch.values())]
res = eval_llm.evaluate(
    data = data,
    checks = [Evals.CONTEXT_RELEVANCE, Evals.FACTUAL_ACCURACY, Evals.RESPONSE_COMPLETENESS]
)

df = pd.DataFrame(res)

# add the langfuse trace_id to the result dataframe
df["trace_id"] = [d['trace_id'] for d in data]
for _, row in df.iterrows():
    for metric_name in ["context_relevance", "factual_accuracy","response_completeness"]:
        langfuse.score(
            name=metric_name,
            value=row["score_"+metric_name],
            trace_id=row["trace_id"]
        )
df.head()

You can visualize these results on Lanfuse:

Tutorial

Open this tutorial in GitHub

Have Questions?

Join our community for any questions or requests

Was this page helpful?

Re-Ranking EvaluationsHelicone