Earlier this year, we released AI functions in public preview, allowing Fabric customers to apply LLM-powered transformations to OneLake data simply and seamlessly, in a single line of code. Since then, we've continued iterating on AI functions in response to your feedback. Let's explore the latest updates, which make AI functions more powerful, more cost-effective, and easier to use than ever before.
As a reminder, AI functions are accessible to all paid Fabric SKUs—and they turbocharge analytics workflows, whether you’re transforming Spark or pandas DataFrames. The demo below shows how AI functions combined with Copilot can accelerate low-code users and experienced developers alike.
Enhanced intelligence—and lower costs—with GPT-4o-mini
We’ve upgraded the default model that powers AI functions to GPT-4o-mini, boosting the feature’s intelligence and effectively reducing its price. The update results in an improved context window (from 16k to 128k input tokens) and significant cost reductions for both input and output tokens. With new optimizations to the system prompts, you can expect better performance and more accurate results from each of the eight AI functions, all while saving Fabric capacity.
No more manual library installations
The library that includes AI functions is now preinstalled on the Fabric 1.3 runtime. That means you can get started with AI-powered data-engineering even faster, skipping the previously required installation steps. The code below shows a simple example of AI-powered translation in PySpark. For more, check out our documentation: Transform and enrich data seamlessly with AI functions.
# This code uses AI. Always review output for mistakes.
# Read terms: https://azure.microsoft.com/support/legal/preview-supplemental-terms/
from synapse.ml.spark.aifunc.DataFrameExtensions import AIFunctions
from synapse.ml.services.openai import OpenAIDefaults
defaults = OpenAIDefaults()
defaults.set_deployment_name("gpt-4o-mini")
df = spark.createDataFrame([
("Hello! How are you doing today?",),
("Tell me what you'd like to know, and I'll do my best to help.",),
("The only thing we have to fear is fear itself.",),
], ["text"])
translations = df.ai.translate(to_lang="spanish", input_col="text", output_col="translations")
display(translations)
AI function support in pure Python notebooks
If you don't need the weight of Spark, you can now use AI functions on pandas DataFrames in Fabric’s pure Python notebooks. The functionality is the same, letting you harness GenAI for data preparation with even more flexibility. Just remember to install the required OpenAI and SynapseML dependencies using the code below.
# Install fixed version of packages
%pip install -q --force-reinstall openai==1.30
# Install latest version of SynapseML-core
%pip install -q --force-reinstall https://mmlspark.blob.core.windows.net/pip/1.0.11-spark3.5/synapseml_core-1.0.11.dev1-py2.py3-none-any.whl
# Install SynapseML-Internal .whl with AI functions library from blob storage:
%pip install -q --force-reinstall https://mmlspark.blob.core.windows.net/pip/1.0.11.1-spark3.5/synapseml_internal-1.0.11.1.dev1-py2.py3-none-any.whl
# This code uses AI. Always review output for mistakes.
# Read terms: https://azure.microsoft.com/support/legal/preview-supplemental-terms/
import synapse.ml.aifunc as aifunc
import pandas as pd
df = pd.DataFrame([
("Scarves"),
("Snow pants"),
("Ski goggles")
], columns=["product"])
df["response"] = df.ai.generate_response("Write a short, punchy email subject line for a winter sale.")
display(df)
Low-code flows directly from your notebook
The syntax for AI functions is so straightforward that you may not need documentation for long. But Fabric notebooks now include a simple interface that generates the code for you. Just select a function, choose an active pandas or Spark DataFrame, and fill in the required inputs.
Next steps
You can learn more about AI functions—and explore starter code samples—using our documentation: Transform and enrich data seamlessly with AI functions. Before getting started, please make note of a few prerequisites:
- The feature requires a paid Fabric SKU, or a P SKU, to use the built-in AI endpoint. You can still use AI functions with a trial capacity if you provide your own Azure OpenAI resource with custom settings.
- Your tenant administrator must enable the tenant switch for Copilot and other features powered by Azure OpenAI. (Depending on your location, you may also need to enable the tenant setting for cross-geo processing.)
We’re eager for you to incorporate AI functions in your own data-science and data-engineering workflows. Let us know what you think by submitting feedback on Fabric Ideas or joining the conversation on the Fabric Community. The features that you request may end up in a future blog post.