🎯 RAGET Testset Generation¶

Waiting to collect data from production to evaluate your RAG agents extensively is a risky business. But building an in-house evaluation dataset is a painful task that requires manual curation and review.

To help with this, the Giskard python library provides RAGET: RAG Evaluation Toolkit, a toolkit to evaluate RAG agents automatically.

ℹ️ You can find a tutorial where we demonstrate the capabilities of RAGET with a simple RAG agent build with LlamaIndex on the IPCC report.

What does RAGET do exactly?¶

RAGET can generate automatically a list of question, reference_answer and reference_context from the knowledge base of the RAG. It relies on a chain of LLM operations to generate realistic questions across different types. You can then use this generated test set to evaluate your RAG agent.

By default, RAGET will create multiple types of questions. Each of them are designed to target and evaluate specific components of the RAG system (for example: the retriever, the generation, or the quality of your knowledge base chunks). During evaluation, RAGET will use a mapping between questions type and RAG components to identify possible weaknesses affecting a specific component of your RAG agent.

RAGET is capable of targeting and evaluating the following components of a RAG agent:

Generator: the LLM used inside the RAG to generate the answers
Retriever: fetch relevant documents from the knowledge base according to a user query
Rewriter (optional): rewrite the user query to make it more relevant to the knowledge base or to account for chat history
Router (optional): filter the query of the user based on his intentions (intentions detection)
Knowledge Base: the set of documents given to the RAG to generate the answers

These are the question types currently supported by RAGET:

Question type	Description	Targeted RAG components
Simple questions	Simple questions generated from an excerpt of the knowledge base Example: What is the capital of France?	`Generator`, `Retriever`, `Router`
Complex questions	Questions made more complex by paraphrasing Example: What is the capital of the country of Victor Hugo?	`Generator`
Distracting questions	Questions made to confuse the retrieval part of the RAG with a distracting element from the knowledge base but irrelevant to the question Example: Italy is beautiful but what is the capital of France?	`Generator`, `Retriever`, `Rewriter`
Situational questions	Questions including user context to evaluate the ability of the generation to produce relevant answer according to the context Example: I am planning a trip to Europe, what is the capital of France?	`Generator`
Double questions	Questions with two distinct parts to evaluate the capabilities of the query rewriter of the RAG Example: What is the capital and the population of France?	`Generator`, `Rewriter`
Conversational questions	Questions made as part of a conversation, first message describes the context of the question that is asked in the last message, also tests the rewriter Example: (two separate messages) I would like to know some information about France. What is its capital city?	`Rewriter`

Before starting¶

Before starting, make sure you have installed the LLM flavor of Giskard:

pip install "giskard[llm]"

To use the RAG test set generation and evaluation tools, you’ll need an API key from the LLM provider that you are using. You can set this API key in your notebook. like this:

OpenAI

import giskard
import os
from giskard.llm.client.openai import OpenAIClient

os.environ["OPENAI_API_KEY"] = "sk-…"

giskard.llm.set_llm_api("openai")
oc = OpenAIClient(model="gpt-4-turbo-preview")
giskard.llm.set_default_client(oc)

Azure OpenAI

Require openai>=1.0.0

import os
from giskard.llm import set_llm_model

os.environ['AZURE_OPENAI_API_KEY'] = '...'
os.environ['AZURE_OPENAI_ENDPOINT'] = 'https://xxx.openai.azure.com'
os.environ['OPENAI_API_VERSION'] = '2023-07-01-preview'


# You'll need to provide the name of the model that you've deployed
# Beware, the model provided must be capable of using function calls
set_llm_model('my-gpt-4-model')

Mistral

import os
from giskard.llm.client.mistral import MistralClient

os.environ["MISTRAL_API_KEY"] = "sk-…"

mc = MistralClient()
giskard.llm.set_default_client(mc)

Ollama

from openai import OpenAI
from giskard.llm.client.openai import OpenAIClient
from giskard.llm.client.mistral import MistralClient

# Setup the Ollama client with API key and base URL
_client = OpenAI(base_url="http://localhost:11434/v1/", api_key="ollama")
oc = OpenAIClient(model="gemma:2b", client=_client)
giskard.llm.set_default_client(oc)

Custom Client

import giskard
from typing import Sequence, Optional
from giskard.llm.client import set_default_client
from giskard.llm.client.base import LLMClient, ChatMessage



class MyLLMClient(LLMClient):
    def __init__(self, my_client):
        self._client = my_client

    def complete(
            self,
            messages: Sequence[ChatMessage],
            temperature: float = 1,
            max_tokens: Optional[int] = None,
            caller_id: Optional[str] = None,
            seed: Optional[int] = None,
            format=None,
    ) -> ChatMessage:
        # Create the prompt
        prompt = ""
        for msg in messages:
            if msg.role.lower() == "assistant":
                prefix = "\n\nAssistant: "
            else:
                prefix = "\n\nHuman: "

            prompt += prefix + msg.content

        prompt += "\n\nAssistant: "

        # Create the body
        params = {
            "prompt": prompt,
            "max_tokens_to_sample": max_tokens or 1000,
            "temperature": temperature,
            "top_p": 0.9,
        }
        body = json.dumps(params)

        response = self._client.invoke_model(
            body=body,
            modelId=self._model_id,
            accept="application/json",
            contentType="application/json",
        )
        data = json.loads(response.get("body").read())

        return ChatMessage(role="assistant", message=data["completion"])

set_default_client(MyLLMClient())

Prepare your Knowledge Base¶

Prepare your data or knowledge base in a pandas DataFrame. Then, create a KnowledgeBase instance with the from_pandas method.

By default, we will use all columns in your data frame to populate your knowledge base. If only certain columns in your dataframe are relevant for the generation of questions, make sure to specify the columns you want to be used with columns argument (see KnowledgeBase).

from giskard.rag import generate_testset, KnowledgeBase

# Load your data and initialize the KnowledgeBase
df = pd.read_csv("path/to/your/knowledge_base.csv")

knowledge_base = KnowledgeBase.from_pandas(df, columns=["column_1", "column_2"])

Generate a test set¶

By default, RAGET automatically generates 6 different question types (these can be selected if needed, see advanced question generation). The total number of questions is divided equally between each question type. To make the question generation more relevant and accurate, you can also provide a description of your agent.

# Generate a testset with 10 questions & answers for each question types (this will take a while)
testset = generate_testset(
    knowledge_base, 
    num_questions=60,
    language='en',  # optional, we'll auto detect if not provided
    agent_description="A customer support chatbot for company X", # helps generating better questions
)

Depending on how many questions you generate, this can take a while. Once you’re done, you can save this generated test set for future use:

# Save the generated testset
testset.save("my_testset.jsonl")

# You can easily load it back
from giskard.rag import QATestset

loaded_testset = QATestset.load("my_testset.jsonl")

You can also convert it to a pandas DataFrame, for quick inspection or further processing:

# Convert it to a pandas dataframe
df = loaded_testset.to_pandas()

Here’s an example of a generated question:

question	reference_context	reference_answer	metadata
For which countries can I track my shipping?	Document 1: We offer free shipping on all orders over $50. For orders below $50, we charge a flat rate of $5.99. We offer shipping services to customers residing in all 50 states of the US, in addition to providing delivery options to Canada and Mexico. Document 2: Once your purchase has been successfully confirmed and shipped, you will receive a confirmation email containing your tracking number. You can simply click on the link provided in the email or visit our website’s order tracking page.	We ship to all 50 states in the US, as well as to Canada and Mexico. We offer tracking for all our shippings.	{“question_type”: “simple”, “seed_document_id”: 1, “topic”: “Shipping policy”}

Each row of the test set contains 5 columns:

question: the generated question
reference_context: the context that can be used to answer the question
reference_answer: the answer to the question (generated with GPT-4)
conversation_history: not shown in the table above, contain the history of the conversation with the agent as a list, only relevant for conversational question, otherwise it contains an empty list.
metadata: a dictionary with various metadata about the question, this includes the question_type, seed_document_id the id of the document used to generate the question and the topic of the question

Advanced configuration of the question generation¶

By default, the test set contains all question types. You can change this by providing question generators to the giskard.rag.generate_testset function. Generators are available inside the question_generators module. For instance to generate only complex and double questions use the following:

from giskard.rag.question_generators import complex_questions, double_questions

testset = generate_testset(
    knowledge_base, 
    question_generators=[complex_questions, double_questions],
)

You can also implement custom question generators, by implementing the interface defined by QuestionGenerator.

What’s next: evaluate your model on the generated testset¶

Once you have generated the test set, you can evaluate your RAG agent using the giskard.rag.evaluate function. Detailed instructions can be found in the RAGET Evaluation section.

Frequently Asked Questions¶

ℹ️ What data are being sent to LLM Providers

In order to perform tasks with language model-assisted detectors, we send the following information to the selected language model provider (e.g., OpenAI, Azure OpenAI, Ollama, Mistral, etc):

Data provided in your knowledge base

Text generated by your model

Model name and description

🌎 Will the test set generation work in any language?

Yes, you can specify the language of the generated questions when calling the generate_testset function. Ultimately, the quality of the generated questions will depend on the LLM performance in the desired language.

Troubleshooting¶

If you encounter any issues, join our Discord community and ask questions in our #support channel.