We propose an approach that perturbs the prompt via masking words to generate several outputs through the LLM and then compares the outputs with the generated output of the unperturbed prompt through vector-based similarity at a semantic level. After this, the word level importance of each masked word in the perturbed sentences is mathematically derived and referred to as the XPLAIN metric.

This repository contains code to run generate the XPLAIN metric for a given input prompt.

Currently models from the family of VERTEX_AI and GOOGLE_AI_STUDIO are supported for LLMs and the Embddeing Models, these include the latest Gemini models aswell.

In order to use the GOOGLE_AI_STUDIO models please ensure that API Key is set as an environment variable.

export GOOGLE_AI_STUDIO_API_KEY="YOUR_API_KEY"

To install the package simply run:

pip install git+https://github.com/dhargopala/xplain.git

examples/google.py contains the code needed to generate the XPLAIN score for the input prompt. This examples utilizes Google AI Studio models.

We first import the package and define the constants.

from xplain.executor import XPLAINMetricCalculator

LLM_MODEL_NAME="gemini-2.5-flash"
EMBEDDER_MODEL_NAME="gemini-embedding-001"

After setting the global values we define the LLM and Embedding model's arguments.

llm_args = {
    "MODEL_NAME": LLM_MODEL_NAME,
    "output_tokens": 65535,   
    "top_k": 40,
    "top_p": 0.95,
    "temperature": 0
}
embedder_args = {
    "MODEL_NAME": EMBEDDER_MODEL_NAME
}

After setting the relevant API arguments, the execution just requires a few lines of code:

xplain_calculator = XPLAINMetricCalculator()

xplain_calculator.select_llm("GOOGLE_AI_STUDIO", **llm_args)
xplain_calculator.select_embedder("GOOGLE_AI_STUDIO", **embedder_args)

scores = xplain_calculator.compute_score("Why do we need sleep?")

Sample Output:

XPLAIN Scores for the prompt: 'Why do we need sleep?' :
-------------------------
Why        | Score: 0.1613
do         | Score: 0.0188
we         | Score: 0.0000
need       | Score: 0.0785
sleep?     | Score: 1.0000
-------------------------