AI-Powered Sentiment Analysis in Microsoft Fabric with Azure OpenAI

[mabdollahi]

Using Microsoft Fabric and Azure OpenAI to Classify Customer Feedback (with PySpark & Power BI)

We’ve all seen how customer feedback holds a goldmine of insights—but going through it manually isn’t scalable. And while AI sounds great, setting it up can feel a bit… intimidating.

The good news? If you're already working with Microsoft Fabric, you're just a few steps away from building your own AI-powered sentiment analysis solution—without needing to train a model or spin up separate services.

In this post (and accompanying video tutorial), I’ll walk you through how to:

✅Store and prepare customer feedback in a Fabric Lakehouse
✅Use PySpark to analyze sentiment using Azure OpenAI
✅Automate the process in a Data Pipeline
✅Visualize results in Power BI

Let’s dive in!

---

🔧What We'll Build

We're going to build an analytical solution that takes in raw customer feedback and automatically classifies each message as Positive, Neutral, or Negative. We'll use:

• Fabric Lakehouse to store the data

• PySpark notebook to process it

• Azure OpenAI (GPT) for the sentiment logic

• Data Pipeline to automate it

• Power BI for the final visualization

All of this is done inside Microsoft Fabric and Azure AI Foundry—no need to leave the ecosystem.

---

🧱 Step 1: Set up your Fabric workspace & lakehouse

Start by creating a new Fabric Workspace and inside it, a Lakehouse.

1. Create Fabric Workspace
   1. Click on "Workspaces" button on the sidebar menu.
   2. Click on "New workspace" button on the opened prompt.
   3. Give it a name and click on "Apply" button. 
2. Create a Lakehouse
   1. Navigate to your worksapce and click on "New item" button. 
   2. Look for "Lakehouse" to filter the items and then click on "Lakehouse". 
   3. Give it a name(CustomerFeedbackAnalysisLH) and click on "Create" button. 
      1. You'll use this Lakehouse to upload your raw feedback data. 
3. Upload the file and create a table out of it.
   1. Navigate to your Lakehouse and click on "Get data > Upload files" item. 
   2. Choose your files and click on "Upload" button. If you don't have a dataset handy, you can grab the one I’ve shared on GitHub (link)—it's a simple CSV with 100 lines of customer feedback. 
   3. You can see the uploaded file in Files folder on your Lakehouse. Choose the file and click on ... button to open the sub-menu. Choose Load to Tables > New Table. 
   4. Give it a name(raw_customer_feedback) and click on "Load" button. 
   5. The data will be shown as a delta table in Tables folder. 

---

🤖Step 2: Set up Azure OpenAI (via Azure AI Studio)

If you haven’t used Azure OpenAI before, here’s what you need to do by follwoing the steps in Create and deploy an Azure OpenAI in Azure AI Foundry Models resource :

1. Head to Azure Portal

2. Create a new Azure Open AI and navigate to AI Foundry

3. Add a deployment of the gpt-35-turbo model

4. Note down your endpoint URL and API key

We’ll use these in the next step to send feedback text to the GPT model and get a sentiment label back.

---

🐍Step 3: Write your PySpark Notebook

Back in Fabric, create a new Notebook using PySpark; in your Lakehouse, click on "New Notebook" from "Open Notebook" menu.

You can start using your codes on Notebook using Phyton or any preferred language.

Load your customer feedback table:

df = spark.read.format("delta").load("Tables/raw_customer_feedback")

Now let’s connect to the OpenAI API and classify each feedback entry. You’ll use Python’s requests module to send text and get a response:

from pyspark.sql.functions import col, udf
import requests

# Azure OpenAI Config
AZURE_OPENAI_ENDPOINT = "https://customname.openai.azure.com/"
AZURE_OPENAI_API_KEY = "secretkey"
DEPLOYMENT_NAME = "gpt-4o"

headers = {
    "Content-Type": "application/json",
    "api-key": AZURE_OPENAI_API_KEY
}

def analyze_sentiment(text):
    try:
        payload = {
            "messages": [{"role": "system", "content": "Analyze sentiment as Positive, Neutral, or Negative"},
                         {"role": "user", "content": text}],
            "max_tokens": 100
        }

        response = requests.post(
            f"{AZURE_OPENAI_ENDPOINT}openai/deployments/{DEPLOYMENT_NAME}/chat/completions?api-version=2023-07-01-preview",
            headers=headers, json=payload
        )

        sentiment = response.json()["choices"][0]["message"]["content"]
        return sentiment
    except Exception as e:
        return "Error"

analyze_sentiment_udf = udf(analyze_sentiment)

Then apply that to each row:

df_enriched = df.withColumn("Sentiment", analyze_sentiment_udf(col("Feedback")))
df_enriched.write.format("delta").mode("overwrite").save("Tables/customer_feedback_enriched")

This adds a new sentiment column to your data and saves it back into your Lakehouse.

You can see the created table with new sentiment field in your Lakehouse.

---

🔄Step 4: Automate with a Data Pipeline

To avoid running the notebook manually every time, create a Data Pipeline.

1. Add a Notebook activity and point it to the notebook you just wrote.

2. Set a trigger to run the pipeline on a schedule, or whenever new data arrives.

3. (Optional) Add a dataflow before the notebook to clean/prepare input data.

Now you’ve got a repeatable pipeline!

---

📊Step 5: Build a Power BI Report

Last step—open Power BI (either within Fabric or using Desktop) and connect to your Lakehouse.

Load the feedback_with_sentiment table and build a few visuals:

• Pie chart showing the sentiment split

• Table of feedback messages with their sentiment

• Time-based trends if you’ve got timestamps

Tip: You can also set custom colors for sentiment categories—like red for Negative, green for Positive—to make things more intuitive.

---

🎥Watch the Full Walkthrough

If you prefer to follow along visually, I’ve recorded a full 25-minute video showing every step, including:

• Creating the workspace & lakehouse

• Writing the PySpark code

• Connecting to Azure OpenAI

• Automating with a pipeline

• Building the report

📂Grab the code and sample data: https://github.com/mehrdadabdollahi/AI-Powered-Sentiment-Analysis-in-Microsoft-Fabric-with-Azure-Ope... 

---

✨Final Thoughts

This project shows how easy it is to plug AI into your data pipelines using Microsoft Fabric—no need to build or train models. With PySpark, Azure OpenAI, and Power BI all in one ecosystem, you get a powerful setup that’s easy to maintain and scale.

If you’re already in the Microsoft data world, this is a great way to dip your toes into AI without leaving the tools you know.

Let me know how it goes—or how you’d extend it further (like adding language detection, tagging, or summarization).

Mehrdad Abdollahi

---