This sample provides a practical demonstration how to perform quality check on LLM results for such tasks as text summarization and translation with Semantic Kernel Filters.

Metrics used in this example:

BERTScore - leverages the pre-trained contextual embeddings from BERT and matches words in candidate and reference sentences by cosine similarity.
BLEU (BiLingual Evaluation Understudy) - evaluates the quality of text which has been machine-translated from one natural language to another.
METEOR (Metric for Evaluation of Translation with Explicit ORdering) - evaluates the similarity between the generated summary and the reference summary, taking into account grammar and semantics.
COMET (Crosslingual Optimized Metric for Evaluation of Translation) - is an open-source framework used to train Machine Translation metrics that achieve high levels of correlation with different types of human judgments.

In this example, SK Filters call dedicated server which is responsible for task evaluation using metrics described above. If evaluation score of specific metric doesn't meet configured threshold, an exception is thrown with evaluation details.

Hugging Face Evaluate Metric library is used to evaluate summarization and translation results.

Prerequisites

Python 3.12
Get Hugging Face API token.
Accept conditions to access Unbabel/wmt22-cometkiwi-da model on Hugging Face portal.

Setup

It's possible to run Python server for task evaluation directly or with Docker.

Run server

Open Python server directory:

cd python-server

Create and active virtual environment:

python -m venv venv
source venv/Scripts/activate # activate on Windows
source venv/bin/activate # activate on Unix/MacOS

Setup Hugging Face API key:

pip install "huggingface_hub[cli]"
huggingface-cli login --token <your_token>

Install dependencies:

pip install -r requirements.txt

Run server:

cd app
uvicorn main:app --port 8080 --reload

Open http://localhost:8080/docs and check available endpoints.

Run server with Docker

Open Python server directory:

cd python-server

Create following Dockerfile:

# syntax=docker/dockerfile:1.2
FROM python:3.12

WORKDIR /code

COPY ./requirements.txt /code/requirements.txt

RUN pip install "huggingface_hub[cli]"
RUN --mount=type=secret,id=hf_token \
    huggingface-cli login --token $(cat /run/secrets/hf_token)

RUN pip install cmake
RUN pip install --no-cache-dir --upgrade -r /code/requirements.txt

COPY ./app /code/app

CMD ["fastapi", "run", "app/main.py", "--port", "80"]

Create .env/hf_token.txt file and put Hugging Face API token in it.
Build image and run container:

docker-compose up --build

Open http://localhost:8080/docs and check available endpoints.

Testing

Open and run QualityCheckWithFilters/Program.cs to experiment with different evaluation metrics, thresholds and input parameters.