If you haven’t already, check out Arun Ulag’s hero blog “FabCon and SQLCon 2026: Unifying databases and Fabric on a single, complete platform” for a complete look at all of our FabCon and SQLCon announcements across both Fabric and our database offerings.
Fabric AI functions let you transform data into insights with a single line of code. Since going generally available, AI functions have helped customers dramatically accelerate their data workflows. We're introducing multimodal support for AI functions (Preview), along with several exciting updates to help you get started faster and manage costs with confidence.
Figure: Fabric notebook using multimodal AI functions to load files into a table, classify and analyze sentiment from photos, and summarize and extract details from PDFs.
Use multimodal AI functions to load files into a table, classify and analyze sentiment from photos, summarize and extract details from PDFs, and more.
What's new?
- Multimodal AI functions (Preview) — Process images, PDFs, and text files directly with AI functions.
- Improved cost transparency — Estimate consumption with a cost calculator, and track usage with a dedicated operation name in the Fabric Capacity Metrics app.
- Eval notebooks — Evaluate and validate AI functions output quality with structured notebook workflows.
- Starter notebooks — Hit the ground running with ready-made notebooks for AI functions.
Multimodal AI functions (Preview)
Until now, AI functions have operated exclusively on text input. With multimodal support, you can now point AI functions directly at images, PDFs, and other files and extract structured insights from them just as easily as you do from text.Supported file types
- Images: JPG, JPEG, PNG, static GIF, WebP
- Documents: PDF
- Text files: MD, TXT, CSV, TSV, JSON, XML, PY, and other text files
New multimodal functions
Multimodal support introduces several new functions alongside compatibility with existing AI functions:aifunc.load(): Automatically reads all files in a folder and generates a structured table. You can optionally pass a `prompt` parameter to guide which information is extracted, and a `schema` parameter to supply a schema for reproducible results.
aifunc.list_file_paths(folder_path): Fetches all valid files from a folder (or glob pattern) and returns them as a series of URLs, ready to use with any AI function.
ai.infer_schema(): Infers a common schema from your files, with or without a guiding prompt. The inferred schema is returned as a list of ExtractLabel objects that can be passed directly to ai.extract().
Use all your favorite AI functions on files
Most AI functions now work with multimodal input. Here are a few examples:Classify insurance claim photos by category:
Summarize PDF Earnings Reports:
Extract invoice date, vendor name, and total amount from invoices:
Analyze sentiment of product review photos:
Use a custom prompt to flag risky clauses in contracts:
Note: passing column_type="path" as a parameter indicates that a column contains file paths rather than plain text.
Improved cost transparency
Understanding and managing costs is critical when working with AI-powered features. We're making it easier to see exactly what AI functions cost with a few improvements:Progress bar with cost calculator
AI functions now include a configurable progress bar that gives you real-time visibility into your job's progress and estimated token usage. You can set the progress bar mode using the use_progress_bar configuration:| Mode | What it shows |
| "basic" (default) | Rows processed and percent completion |
| "stats" | Everything in basic, plus estimated total input tokens, predicted output tokens, and a running total CU prediction that updates in real time — with a final total CU count when the job completes |
| "disable" | Turns the progress bar off entirely |
In stats mode, you get a running total CU prediction that updates in real time as rows are processed, and a final CU total when the job completes. This is a game changer for cost management. You know exactly what to expect in terms of capacity consumption while your AI function is still running, giving you the confidence to scale up or the signal to adjust before the job finishes.
Learn more about the progress bar cost calculator in the custom configurations article.
Figure: Sample progress bar with the "basic", "stats", and "disable" modes.
Dedicated operation name in Fabric Capacity Metrics
AI functions now have their own dedicated AI Functions operation in the Fabric Capacity Metrics App. Previously, AI functions usage was reported under Spark or Dataflows Gen2 operations, making it difficult to isolate. Now, AI functions consumption appears as its own line item—whether you call them from a Notebook or Dataflows Gen2—giving you clear visibility into how much capacity AI functions are using.For full details, see our recent blog post: Billing updates: new operations for Fabric AI Functions and AI Services.
Eval Notebooks
How do you know if the AI-generated results are good enough for your use case? We're introducing evaluation notebooks, a set of Python notebooks hosted in the Fabric samples GitHub repo that helps you systematically evaluate the output quality of AI functions using LLM-as-a-Judge, a technique where a larger frontier model acts as an evaluator to assess quality.How it works
Each eval notebook walks you through a consistent workflow:- Run an AI function on your data using an executor model (e.g., gpt-4.1-mini).
- Evaluate each output using a judge model (e.g., gpt-5 with reasoning) to generate ground truth labels or quality scores.
- Calculate metrics such as accuracy, precision, recall, F1 score, coherence, consistency, relevance, and more, where relevant.
- Identify which predictions need review and visualize the results.
- Refine, optionally adjust labels, prompts, or model configurations and re-evaluate.
Use it as a template
The eval notebooks are designed as starting points. You may replace the sample data, adapt the evaluation prompts and criteria to your use case, and build repeatable quality checks that you can run as your data or requirements change. The notebooks also include troubleshooting guidance and options for improving quality, such as switching to a larger model or using ai.generate_response with custom prompts for full control.Starter notebooks
Getting started with AI functions is now easier than ever. We're releasing starter notebooks that walk you through an end-to-end AI functions workflow—from setup and configuration to running AI functions on your data. These notebooks provide a copy-and-customize starting point for your own projects.Download the starter notebook and import it into your Fabric workspace to start transforming your data with AI in minutes.
Next Steps
- Try multimodal AI functions: point to your images, PDFs, or text files, and start extracting insights with a single line of code.
- Explore the documentation: learn more about AI functions and multimodal capabilities.
- Download the starter notebooks and eval notebooks.
- Share your feedback: Submit ideas on Fabric Ideas and join the conversation in the Fabric Community.