Groundedness Evaluator an Evaluator that assesses whether answers are accurate based on the given context. This Evaluator can be used to evaluate hallucinations in RAG's responses. In this tutorial, we will look at how to evaluate Groundedness using Upstage Groundedness Checker (UpstageGroundednessCheck) and a custom-made Groundedness Checker.
Setting up your environment is the first step. See the Environment Setup guide for more details.
[Note]
The langchain-opentutorial is a package of easy-to-use environment setup guidance, useful functions and utilities for tutorials. Check out the langchain-opentutorial for more details.
You can alternatively set API keys such as OPENAI_API_KEY in a .env file and load them.
[Note] This is not necessary if you've already set the required API keys in previous steps.
# Load API keys from .env file
from dotenv import load_dotenv
load_dotenv(override=True)
True
Define a function for RAG performance testing
Let's create an RAG system that will be used for testing.
from typing import Callable
from langchain_community.document_loaders import PyMuPDFLoader
from langchain_community.vectorstores import FAISS
from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import PromptTemplate
from langchain_core.runnables import RunnablePassthrough
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_text_splitters import RecursiveCharacterTextSplitter
def ask_question_with_llm(llm: ChatOpenAI, file_path: str) -> Callable[[dict], dict]:
"""Create a function that answers questions.
Args:
llm (ChatOpenAI): Language Model
file_path (str): Path to the PDF file
Returns:
(Callable[[dict], dict]): A function that answers questions
"""
# Load documents
loader = PyMuPDFLoader(file_path)
docs = loader.load()
# Split given documents
text_splitter = RecursiveCharacterTextSplitter(chunk_size=300, chunk_overlap=50)
split_documents = text_splitter.split_documents(docs)
# Create a retriever
embeddings = OpenAIEmbeddings(model="text-embedding-3-small")
vectorstore = FAISS.from_documents(documents=split_documents, embedding=embeddings)
retriever = vectorstore.as_retriever()
# Create a chain
prompt = PromptTemplate.from_template(
"""
You are an assistant for question-answering tasks.
Use the following pieces of retrieved context to answer the question.
If you don't know the answer, just say that you don't know.
#Context:
{context}
#Question:
{question}
#Answer:
"""
)
rag_chain = (
{
"context": retriever,
"question": RunnablePassthrough(),
}
| prompt
| llm
| StrOutputParser()
)
def _ask_question(inputs: dict) -> dict:
# Search for context related to the question
context = retriever.invoke(inputs["question"])
# Combine the retrieved documents into a single string
context = "\n".join([doc.page_content for doc in context])
# Return a dictionary containing the question, context, and answer
return {
"question": inputs["question"],
"context": context,
"answer": rag_chain.invoke(inputs["question"]),
}
return _ask_question
gpt_chain = ask_question_with_llm(
llm=ChatOpenAI(model="gpt-4o-mini", temperature=0),
file_path="data/Newwhitepaper_Agents2.pdf",
)
Set Groundedness Checkers
To evaluate groundedness, UpstageGroundednessCheck and Custom Groundedness Checker will be used.
Set UpstageGroundednessCheck
To use Upstage's Groundedness Checker (UpstageGroundednessCheck), you need to obtain an API key from the link below.
# Run Groundedness Checker for evaluation
request_input = {
"context": "Teddy's gender is male and he operates the Teddynote YouTube channel.",
"answer": "Teddy is a male.",
}
response = upstage_groundedness_check.invoke(request_input)
print(response)
grounded
# Run Groundedness Checker for evaluation
request_input = {
"context": "Teddy's gender is male and he operates the Teddynote YouTube channel.",
"answer": "Teddy is a female.",
}
response = upstage_groundedness_check.invoke(request_input)
print(response)
notGrounded
Define UpstageGroundednessCheck Evaluator. It will be used in evaluate function.
Create a custom Groundedness Checker using OpenAI's model.
For this tutorial, we'll use the target retrieval-answer. If you want to use other targets ('question-answer' or 'question-retrieval'), you should change the description in GroundednessScore and the prompt template in GroundednessChecker.create() accordingly.
from langchain_core.runnables import RunnableSequence
from langchain_openai import ChatOpenAI
from langsmith.schemas import Example, Run
from pydantic import BaseModel, Field
class GroundednessScore(BaseModel):
"""Binary scores for relevance checks"""
score: str = Field(
description="relevant or not relevant. Answer 'yes' if the answer is relevant to the retrieved document else answer 'no'"
)
class GroundednessChecker:
"""This class is used to evaluate the accuracy of a document.
It returns 'yes' or 'no' as the evaluation result.
Attributes:
llm (ChatOpenAI): Language Model instance.
"""
def __init__(self, llm: ChatOpenAI) -> None:
self.llm = llm
def create(self) -> RunnableSequence:
"""
Create a chain for evaluating accuracy.
Returns:
Chain: A chain object that can evaluate accuracy
"""
llm = self.llm.with_structured_output(GroundednessScore)
# 프롬프트 선택
template = """You are a grader assessing relevance of a retrieved document to a user question. \n
Here is the retrieved document: \n\n {context} \n\n
Here is the answer: {answer} \n
If the document contains keyword(s) or semantic meaning related to the user answer, grade it as relevant. \n
Give a binary score 'yes' or 'no' score to indicate whether the retrieved document is relevant to the answer.
"""
input_vars = ["context", "answer"]
# Create a prompt
prompt = PromptTemplate(
template=template,
input_variables=input_vars,
)
# Create a chain
chain = prompt | llm
return chain
# Create a Groundedness Checker
custom_groundedness_check = GroundednessChecker(
ChatOpenAI(model="gpt-4o-mini", temperature=0)
).create()
def custom_groundedness_check_evaluator(run: Run, example: Example) -> dict:
# Get generated answer and context
answer = run.outputs.get("answer", "")
context = run.outputs.get("context", "")
# Groundedness Check
groundedness_score = custom_groundedness_check.invoke(
{"answer": answer, "context": context}
)
groundedness_score = groundedness_score.score == "yes"
return {"key": "groundedness_score", "score": int(groundedness_score)}
Evaluate Groundedness using Upstage's and Custom Groundedness Checker
Evaluate Groundedness using Upstage's and Custom Groundedness Checker.
Before this, check if the dataset exists created before. If you don't have the dataset, create one referring to 04-LangSmith-Dataset.
In this tutorial, we'll use custom Q&A dataset referred to Google Whitepaper on AI Agents .
from langsmith.client import Client
client = Client()
datasets = client.list_datasets()
for dataset in datasets:
print(dataset.name)
for example in client.list_examples(dataset_name=dataset.name):
print("inputs :", example.dict()["inputs"]["question"])
print("outputs :", example.dict()["outputs"]["answer"])
print("--------------------------------")
RAG_EVAL_DATASET
inputs : What are the three targeted learnings to enhance model performance?
outputs : The three targeted learning approaches to enhance model performance mentioned in the context are:
1. In-context learning: This involves providing a generalized model with a prompt, tools, and few-shot examples at inference time, allowing it to learn "on the fly."
2. Fine-tuning based learning: This method involves training a model using a larger dataset of specific examples prior to inference, helping the model understand when and how to apply certain tools.
3. Using external memory: This includes examples like the 'Example Store' in Vertex AI extensions or data stores in RAG-based architecture, which help models access specific knowledge.
--------------------------------
inputs : What are the key functions of an agent's orchestration layer in achieving goals?
outputs : The key functions of an agent's orchestration layer in achieving goals include structuring reasoning, planning, and decision-making, as well as guiding the agent's actions. It involves taking in information, performing internal reasoning, and generating informed decisions or responses. The orchestration layer can utilize various reasoning techniques such as ReAct, Chain-of-Thought, and Tree-of-Thoughts to facilitate these processes. Additionally, it can leverage language models and external tools to transition through states and execute complex tasks autonomously.
--------------------------------
inputs : List up the name of the authors
outputs : The authors are Julia Wiesinger, Patrick Marlow, and Vladimir Vuskovic.
--------------------------------
inputs : What is Tree-of-thoughts?
outputs : Tree-of-thoughts (ToT) is a prompt engineering framework that is well suited for exploration or strategic lookahead tasks. It generalizes over chain-of-thought prompting and allows the model to explore various thought chains that serve as intermediate steps for general problem solving with language models.
--------------------------------
inputs : What is the framework used for reasoning and planning in agent?
outputs : The frameworks used for reasoning and planning in agents include ReAct, Chain-of-Thought, and Tree-of-Thoughts. These frameworks provide a structure for the orchestration layer to perform internal reasoning and generate informed decisions or responses.
--------------------------------
inputs : How do agents differ from standalone language models?
outputs : Agents can use tools to access real-time data and perform actions, whereas models rely solely on their training data.
Agents maintain session history for multi-turn reasoning, while models lack native session management.
Agents incorporate cognitive architectures for advanced reasoning, such as ReAct or Chain-of-Thought.
--------------------------------
from langsmith.evaluation import evaluate
dataset_name = "RAG_EVAL_DATASET"
# Run evaluation
experiment_results = evaluate(
gpt_chain,
data=dataset_name,
evaluators=[upstage_groundedness_check_evaluator, custom_groundedness_check_evaluator],
experiment_prefix="GROUNDEDNESS_EVAL",
metadata={
"variant": "Hallucination evaluation using Upstage's and Custom Groundedness Checker",
},
)
View the evaluation results for experiment: 'GROUNDEDNESS_EVAL-929f91f7' at:
https://smith.langchain.com/o/9089d1d3-e786-4000-8468-66153f05444b/datasets/9b4ca107-33fe-4c71-bb7f-488272d895a3/compare?selectedSessions=fd7bec12-fc0d-41d9-bf21-3db571a54f15
0it [00:00, ?it/s]
Comprehensive evaluation of dataset using summary evaluators
This is useful when you want to evaluate the entire dataset. (The previous step was for individual data evaluation.)
from typing import List
from langsmith.schemas import Example, Run
def upstage_groundedness_check_summary_evaluator(
runs: List[Run], examples: List[Example]
) -> dict:
def is_grounded(run: Run) -> bool:
context = run.outputs["context"]
answer = run.outputs["answer"]
return (
upstage_groundedness_check.invoke({"context": context, "answer": answer})
== "grounded"
)
groundedness_scores = sum(1 for run in runs if is_grounded(run))
return {"key": "groundedness_score", "score": groundedness_scores / len(runs)}
def custom_groundedness_check_summary_evaluator(
runs: List[Run], examples: List[Example]
) -> dict:
def is_grounded(run: Run) -> bool:
context = run.outputs["context"]
answer = run.outputs["answer"]
return (
custom_groundedness_check.invoke({"context": context, "answer": answer}).score
== "yes"
)
groundedness_scores = sum(1 for run in runs if is_grounded(run))
return {"key": "groundedness_score", "score": groundedness_scores / len(runs)}
from langsmith.evaluation import evaluate
# Run evaluation
experiment_result1 = evaluate(
gpt_chain,
data=dataset_name,
summary_evaluators=[
upstage_groundedness_check_summary_evaluator,
],
experiment_prefix="GROUNDEDNESS_UPSTAGE_SUMMARY_EVAL",
# Set experiment metadata
metadata={
"variant": "Hallucination evaluation using Upstage Groundedness Checker",
},
)
# Run evaluation
experiment_result2 = evaluate(
gpt_chain,
data=dataset_name,
summary_evaluators=[
custom_groundedness_check_summary_evaluator,
],
experiment_prefix="GROUNDEDNESS_CUSTOM_SUMMARY_EVAL",
# Set experiment metadata
metadata={
"variant": "Hallucination evaluation using Custom Groundedness Checker",
},
)
View the evaluation results for experiment: 'GROUNDEDNESS_UPSTAGE_SUMMARY_EVAL-6d673c3a' at:
https://smith.langchain.com/o/9089d1d3-e786-4000-8468-66153f05444b/datasets/9b4ca107-33fe-4c71-bb7f-488272d895a3/compare?selectedSessions=67a47286-2224-48d3-94d6-e2bb87a47aef
0it [00:00, ?it/s]
View the evaluation results for experiment: 'GROUNDEDNESS_CUSTOM_SUMMARY_EVAL-af449083' at:
https://smith.langchain.com/o/9089d1d3-e786-4000-8468-66153f05444b/datasets/9b4ca107-33fe-4c71-bb7f-488272d895a3/compare?selectedSessions=72ae6cbd-e53b-4311-9c2a-654585da9c5b
0it [00:00, ?it/s]
