aidial-rag-eval 0.6.0.dev23

Library for RAG evaluation

AI DIAL RAG EVAL

Overview

Library designed for RAG (Retrieval-Augmented Generation) evaluation, where retrieval and generation metrics are calculated.

Usage

Install the library using pip:

pip install aidial-rag-eval

spaCy language model

The generation metrics require the English language model for spaCy. Download it after installation:

python -m spacy download en_core_web_sm

Alternatively, you can install the model directly via URL:

pip install https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.8.0/en_core_web_sm-3.8.0-py3-none-any.whl

Or as a Poetry dependency:

en-core-web-sm = {url = "https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.8.0/en_core_web_sm-3.8.0-py3-none-any.whl"}

Example

The example of how to get retrieval metrics along with answer inference based on the context.

import pandas as pd
from langchain_openai import AzureChatOpenAI
from aidial_rag_eval import create_rag_eval_metrics_report
from aidial_rag_eval.metric_binds import CONTEXT_TO_ANSWER_INFERENCE,\
    ANSWER_TO_GROUND_TRUTH_INFERENCE, GROUND_TRUTH_TO_ANSWER_INFERENCE,\
    ANSWER_TO_FACTS_INFERENCE, FACTS_TO_ANSWER_INFERENCE

llm = AzureChatOpenAI(model="gemini-2.5-flash-lite")

df_ground_truth = pd.DataFrame([
    {
        "question": "What is the diameter of the Earth and the name of the biggest ocean?",
        "documents": ["earth.pdf"],
        "facts": ["The diameter of the Earth is approximately 12,742 kilometers.", "The biggest ocean on Earth is the Pacific Ocean."],
        "answer": "The Earth's diameter measures about 12,742 kilometers, and the Pacific Ocean is the largest ocean on our planet."
    },])
df_answer = pd.DataFrame([
    {
        "question": "What is the diameter of the Earth and the name of the biggest ocean?",
        "documents": ["earth.pdf"],
        "context":  [
            "The Earth, our home planet, is the third planet from the sun. It's the only planet known to have an atmosphere containing free oxygen and oceans of liquid water on its surface. The diameter of the Earth is approximately 12,742 kilometers.",
            "The Pacific Ocean is the largest and deepest of Earth's oceanic divisions, extending from the Arctic Ocean in the north to the Southern Ocean in the south."
        ],
        "answer": "The Earth has a diameter of approximately 12,742 kilometers."
    },
])

df_metrics = create_rag_eval_metrics_report(
    df_ground_truth,
    df_answer,
    llm=llm,
    metric_binds=[
        CONTEXT_TO_ANSWER_INFERENCE,
        ANSWER_TO_GROUND_TRUTH_INFERENCE,
        GROUND_TRUTH_TO_ANSWER_INFERENCE,
        ANSWER_TO_FACTS_INFERENCE,
        FACTS_TO_ANSWER_INFERENCE,
    ],
)
print(df_metrics[["facts_ranks", "recall", 'precision', 'mrr', 'f1', 'ctx_ans_inference', 'ans_gt_inference', 'gt_ans_inference', 'ans_fct_inference', 'fct_ans_inference']])

It is expected to see the following results:

recall	precision	mrr	f1	ctx_ans_inference	ans_gt_inference	gt_ans_inference	ans_fct_inference	fct_ans_inference
0.5	0.5	0.5	0.5	1.0	0.5	1.0	0.5	1.0

In this table:

• "recall" of 0.5 indicates that only 1 out of 2 ground truth facts were found in the context.
• "precision" of 0.5 reflects that just 1 context chunk out of 2 includes any ground truth facts.
• The prefix of the inference metrics signifies the premise and hypothesis in the following format: premise_hypothesis_inference.
  • "ctx" refers to 'context'
  • "ans" refers to 'answer'
  • "gt" refers to 'ground truth answer'
  • "fct" refers to 'facts'
• "ctx_ans_inference" of 1.0 means our answer can be fully derived from the context.
• "ans_gt_inference" of 0.5 means the ground truth answer is only partially entailed by our answer.
• "gt_ans_inference" of 1.0 means our answer can be fully derived from the ground truth answer.
• "ans_fct_inference" of 0.5 means only half of the ground truth facts are entailed by our answer (the Pacific Ocean fact is missing).
• "fct_ans_inference" of 1.0 means our answer can be fully derived from the ground truth facts.

Recommended models

The algorithm is token-intensive. Considering the balance between quality and price, the following models are recommended:

• gemini-3.1-flash-lite
• gemini-2.5-flash-lite
• gpt-5-mini
• gpt-5-nano
• gpt-5.4-mini

Developer environment

This project uses Python>=3.11 and Poetry>=2.2.1 as a dependency manager.

Check out Poetry's documentation on how to install it on your system before proceeding.

To install requirements:

poetry install

This will install all requirements for running the package, linting, formatting and tests.

Lint

Run the linting before committing:

make lint

To auto-fix formatting issues run:

make format

Test

Run unit tests locally for available python versions:

make test

Run unit tests for the specific python version:

make test PYTHON=3.11

The generation evaluation requires an access to the LLM. The generation evaluation tests (located in tests/llm_tests directory) use cached LLM responses by default. To run the tests with real LLM responses, you need add --llm-mode=real argument to the test command:

make test PYTHON=3.11 ARGS="--llm-mode=real"

The test run with real LLM responses requires the following environment variables to be set:

Variable	Description
DIAL_URL	The URL of the DIAL server.
DIAL_API_KEY	The API key for the DIAL server.

Copy .env.example to .env and customize it for your environment.

Clean

To remove the virtual environment and build artifacts run:

make clean

Build

To build the package run:

make build