Evaluation with DeepEval

This guide demonstrates how to use DeepEval to evaluate a Retrieval-Augmented Generation (RAG) pipeline built upon Milvus.

The RAG system combines a retrieval system with a generative model to generate new text based on a given prompt. The system first retrieves relevant documents from a corpus using Milvus, and then uses a generative model to generate new text based on the retrieved documents.

DeepEval is a framework that helps you evaluate your RAG pipelines. There are existing tools and frameworks that help you build these pipelines but evaluating it and quantifying your pipeline performance can be hard. This is where DeepEval comes in.

Prerequisites

Before running this notebook, make sure you have the following dependencies installed:

$ pip install --upgrade pymilvus openai requests tqdm pandas deepeval

We will use OpenAI as the LLM in this example. You should prepare the api key OPENAI_API_KEY as an environment variable.

import os

os.environ["OPENAI_API_KEY"] = "sk-*****************"

Define the RAG pipeline

We will define the RAG class that use Milvus as the vector store, and OpenAI as the LLM. The class contains the load method, which loads the text data into Milvus, the retrieve method, which retrieves the most similar text data to the given question, and the answer method, which answers the given question with the retrieved knowledge.

from typing import List
from tqdm import tqdm
from openai import OpenAI
from pymilvus import MilvusClient


class RAG:
    """
    RAG(Retrieval-Augmented Generation) class built upon OpenAI and Milvus.
    """

    def __init__(self, openai_client: OpenAI, milvus_client: MilvusClient):
        self._prepare_openai(openai_client)
        self._prepare_milvus(milvus_client)

    def _emb_text(self, text: str) -> List[float]:
        return (
            self.openai_client.embeddings.create(input=text, model=self.embedding_model)
            .data[0]
            .embedding
        )

    def _prepare_openai(
        self,
        openai_client: OpenAI,
        embedding_model: str = "text-embedding-3-small",
        llm_model: str = "gpt-4o-mini",
    ):
        self.openai_client = openai_client
        self.embedding_model = embedding_model
        self.llm_model = llm_model
        self.SYSTEM_PROMPT = """
            Human: You are an AI assistant. You are able to find answers to the questions from the contextual passage snippets provided.
        """
        self.USER_PROMPT = """
            Use the following pieces of information enclosed in <context> tags to provide an answer to the question enclosed in <question> tags.
            <context>
            {context}
            </context>
            <question>
            {question}
            </question>
        """

    def _prepare_milvus(
        self, milvus_client: MilvusClient, collection_name: str = "rag_collection"
    ):
        self.milvus_client = milvus_client
        self.collection_name = collection_name
        if self.milvus_client.has_collection(self.collection_name):
            self.milvus_client.drop_collection(self.collection_name)
        embedding_dim = len(self._emb_text("demo"))
        self.milvus_client.create_collection(
            collection_name=self.collection_name,
            dimension=embedding_dim,
            metric_type="IP",
            consistency_level="Strong",
        )

    def load(self, texts: List[str]):
        """
        Load the text data into Milvus.
        """
        data = []
        for i, line in enumerate(tqdm(texts, desc="Creating embeddings")):
            data.append({"id": i, "vector": self._emb_text(line), "text": line})
        self.milvus_client.insert(collection_name=self.collection_name, data=data)

    def retrieve(self, question: str, top_k: int = 3) -> List[str]:
        """
        Retrieve the most similar text data to the given question.
        """
        search_res = self.milvus_client.search(
            collection_name=self.collection_name,
            data=[self._emb_text(question)],
            limit=top_k,
            search_params={"metric_type": "IP", "params": {}},  # inner product distance
            output_fields=["text"],  # Return the text field
        )
        retrieved_texts = [res["entity"]["text"] for res in search_res[0]]
        return retrieved_texts[:top_k]

    def answer(
        self,
        question: str,
        retrieval_top_k: int = 3,
        return_retrieved_text: bool = False,
    ):
        """
        Answer the given question with the retrieved knowledge.
        """
        retrieved_texts = self.retrieve(question, top_k=retrieval_top_k)
        user_prompt = self.USER_PROMPT.format(
            context="\n".join(retrieved_texts), question=question
        )
        response = self.openai_client.chat.completions.create(
            model=self.llm_model,
            messages=[
                {"role": "system", "content": self.SYSTEM_PROMPT},
                {"role": "user", "content": user_prompt},
            ],
        )
        if not return_retrieved_text:
            return response.choices[0].message.content
        else:
            return response.choices[0].message.content, retrieved_texts

Let’s initialize the RAG class with OpenAI and Milvus clients.

openai_client = OpenAI()
milvus_client = MilvusClient(uri="./milvus_demo.db")

my_rag = RAG(openai_client=openai_client, milvus_client=milvus_client)

Run the RAG pipeline and get results

We use the Milvus development guide to be as the private knowledge in our RAG, which is a good data source for a simple RAG pipeline.

Download it and load it into the rag pipeline.

import urllib.request
import os

url = "https://raw.githubusercontent.com/milvus-io/milvus/master/DEVELOPMENT.md"
file_path = "./Milvus_DEVELOPMENT.md"

if not os.path.exists(file_path):
    urllib.request.urlretrieve(url, file_path)
with open(file_path, "r") as file:
    file_text = file.read()

text_lines = file_text.split("# ")
my_rag.load(text_lines)

Creating embeddings: 100%|██████████| 47/47 [00:20<00:00,  2.26it/s]

Let’s define a query question about the content of the development guide documentation. And then use the answer method to get the answer and the retrieved context texts.

question = "what is the hardware requirements specification if I want to build Milvus and run from source code?"
my_rag.answer(question, return_retrieved_text=True)

('The hardware requirements specification to build and run Milvus from source code is as follows:\n\n- 8GB of RAM\n- 50GB of free disk space',
 ['Hardware Requirements\n\nThe following specification (either physical or virtual machine resources) is recommended for Milvus to build and run from source code.\n\n```\n- 8GB of RAM\n- 50GB of free disk space\n```\n\n##',
  'Building Milvus on a local OS/shell environment\n\nThe details below outline the hardware and software requirements for building on Linux and MacOS.\n\n##',
  "Software Requirements\n\nAll Linux distributions are available for Milvus development. However a majority of our contributor worked with Ubuntu or CentOS systems, with a small portion of Mac (both x86_64 and Apple Silicon) contributors. If you would like Milvus to build and run on other distributions, you are more than welcome to file an issue and contribute!\n\nHere's a list of verified OS types where Milvus can successfully build and run:\n\n- Debian/Ubuntu\n- Amazon Linux\n- MacOS (x86_64)\n- MacOS (Apple Silicon)\n\n##"])

Now let’s prepare some questions with its corresponding ground truth answers. We get answers and contexts from our RAG pipeline.

from datasets import Dataset
import pandas as pd

question_list = [
    "what is the hardware requirements specification if I want to build Milvus and run from source code?",
    "What is the programming language used to write Knowhere?",
    "What should be ensured before running code coverage?",
]
ground_truth_list = [
    "If you want to build Milvus and run from source code, the recommended hardware requirements specification is:\n\n- 8GB of RAM\n- 50GB of free disk space.",
    "The programming language used to write Knowhere is C++.",
    "Before running code coverage, you should make sure that your code changes are covered by unit tests.",
]
contexts_list = []
answer_list = []
for question in tqdm(question_list, desc="Answering questions"):
    answer, contexts = my_rag.answer(question, return_retrieved_text=True)
    contexts_list.append(contexts)
    answer_list.append(answer)

df = pd.DataFrame(
    {
        "question": question_list,
        "contexts": contexts_list,
        "answer": answer_list,
        "ground_truth": ground_truth_list,
    }
)
rag_results = Dataset.from_pandas(df)
df

/Users/eureka/miniconda3/envs/zilliz/lib/python3.9/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html
  from .autonotebook import tqdm as notebook_tqdm
Answering questions: 100%|██████████| 3/3 [00:03<00:00,  1.06s/it]

Evaluating Retriever

When evaluating a retriever in large language model (LLM) systems, it’s crucial to assess the following:

1. Ranking Relevance: How effectively the retriever prioritizes relevant information over irrelevant data.

2. Contextual Retrieval: The ability to capture and retrieve contextually relevant information based on the input.

3. Balance: How well the retriever manages text chunk size and retrieval scope to minimize irrelevancies.

Together, these factors provide a comprehensive understanding of how the retriever prioritizes, captures, and presents the most useful information.

from deepeval.metrics import (
    ContextualPrecisionMetric,
    ContextualRecallMetric,
    ContextualRelevancyMetric,
)
from deepeval.test_case import LLMTestCase
from deepeval import evaluate

contextual_precision = ContextualPrecisionMetric()
contextual_recall = ContextualRecallMetric()
contextual_relevancy = ContextualRelevancyMetric()

test_cases = []

for index, row in df.iterrows():
    test_case = LLMTestCase(
        input=row["question"],
        actual_output=row["answer"],
        expected_output=row["ground_truth"],
        retrieval_context=row["contexts"],
    )
    test_cases.append(test_case)

# test_cases
result = evaluate(
    test_cases=test_cases,
    metrics=[contextual_precision, contextual_recall, contextual_relevancy],
    print_results=False,  # Change to True to see detailed metric results
)

/Users/eureka/miniconda3/envs/zilliz/lib/python3.9/site-packages/deepeval/__init__.py:49: UserWarning: You are using deepeval version 1.1.6, however version 1.2.2 is available. You should consider upgrading via the "pip install --upgrade deepeval" command.
  warnings.warn(

✨ You're running DeepEval's latest Contextual Precision Metric! (using gpt-4o, strict=False, async_mode=True)...

✨ You're running DeepEval's latest Contextual Recall Metric! (using gpt-4o, strict=False, async_mode=True)...

✨ You're running DeepEval's latest Contextual Relevancy Metric! (using gpt-4o, strict=False, async_mode=True)...

Event loop is already running. Applying nest_asyncio patch to allow async execution...


Evaluating 3 test case(s) in parallel: |██████████|100% (3/3) [Time Taken: 00:11,  3.91s/test case]

✓ Tests finished 🎉! Run 'deepeval login' to view evaluation results on Confident AI. 
‼️  NOTE: You can also run evaluations on ALL of deepeval's metrics directly on Confident AI instead.

Evaluating Generation

To assess the quality of generated outputs in large language models (LLMs), it’s important to focus on two key aspects:

1. Relevance: Evaluate whether the prompt effectively guides the LLM to generate helpful and contextually appropriate responses.

2. Faithfulness: Measure the accuracy of the output, ensuring the model produces information that is factually correct and free from hallucinations or contradictions. The generated content should align with the factual information provided in the retrieval context.

These factors together ensure that the outputs are both relevant and reliable.

from deepeval.metrics import AnswerRelevancyMetric, FaithfulnessMetric
from deepeval.test_case import LLMTestCase
from deepeval import evaluate

answer_relevancy = AnswerRelevancyMetric()
faithfulness = FaithfulnessMetric()

test_cases = []

for index, row in df.iterrows():
    test_case = LLMTestCase(
        input=row["question"],
        actual_output=row["answer"],
        expected_output=row["ground_truth"],
        retrieval_context=row["contexts"],
    )
    test_cases.append(test_case)

# test_cases
result = evaluate(
    test_cases=test_cases,
    metrics=[answer_relevancy, faithfulness],
    print_results=False,  # Change to True to see detailed metric results
)

✨ You're running DeepEval's latest Answer Relevancy Metric! (using gpt-4o, strict=False, async_mode=True)...

✨ You're running DeepEval's latest Faithfulness Metric! (using gpt-4o, strict=False, async_mode=True)...

Event loop is already running. Applying nest_asyncio patch to allow async execution...


Evaluating 3 test case(s) in parallel: |██████████|100% (3/3) [Time Taken: 00:11,  3.97s/test case]

✓ Tests finished 🎉! Run 'deepeval login' to view evaluation results on Confident AI. 
‼️  NOTE: You can also run evaluations on ALL of deepeval's metrics directly on Confident AI instead.