Edit

Share via

Foundry Tools in Fabric (preview)

Important

This feature is in preview.

Foundry Tools help developers and organizations rapidly create intelligent, cutting-edge, market-ready, and responsible applications with prebuilt and customizable APIs and models. Formerly named Azure Cognitive Services, Foundry Tools empower developers even when they don't have direct AI or data science skills or knowledge. The goal of Foundry Tools is to help developers create applications that can see, hear, speak, understand, and even begin to reason.

Fabric provides two options to use Foundry Tools:

  • Pre-built AI models in Fabric (preview)

    Fabric seamlessly integrates with Foundry Tools, allowing you to enrich your data with prebuilt AI models without any prerequisite. We recommend this option because you can use your Fabric authentication to access Foundry Tools, and all usages are billed against your Fabric capacity. This option is currently in public preview, with limited Foundry Tools available.

    Fabric offers Azure OpenAI Service, Text Analytics, and Azure Translator in Foundry Tools by default, with support for both SynapseML and the RESTful API. You can also use the OpenAI Python Library to access Azure OpenAI service in Fabric. For more information about available models, visit prebuilt AI models in Fabric.

  • Bring your own key (BYOK)

    You can provision your Foundry Tools on Azure, and bring your own key to use them from Fabric. If the prebuilt AI models don't yet support the desired Foundry Tools, you can still use BYOK (Bring your own key).

    To learn more about how to use Foundry Tools with BYOK, visit Foundry Tools in SynapseML with bring your own key.

Prebuilt AI models in Fabric (preview)

Azure OpenAI Service

REST API, Python SDK, SynapseML, AI Functions

Text Analytics

REST API, SynapseML

  • Language detection: detects language of the input text
  • Sentiment analysis: returns a score between 0 and 1, to indicate the sentiment in the input text
  • Key phrase extraction: identifies the key talking points in the input text
  • Personally Identifiable Information(PII) entity recognition: identify, categorize, and redact sensitive information in the input text
  • Named entity recognition: identifies known entities and general named entities in the input text
  • Entity linking: identifies and disambiguates the identity of entities found in text

Translator

REST API, SynapseML

  • Translate: Translates text
  • Transliterate: Converts text in one language, in one script, to another script.

Available regions

Available regions for Azure OpenAI Service

For the list of Azure regions where prebuilt Foundry Tools in Fabric are now available, visit the Available regions section of the Overview of Copilot in Fabric and Power BI (preview) article.

Available regions for Text Analytics and Translator

Prebuilt Text Analytics and the Translator in Fabric are now available for public preview in the Azure regions listed in this article. If you don't find your Microsoft Fabric home region in this article, you can still create a Microsoft Fabric capacity in a supported region. For more information, visit Buy a Microsoft Fabric subscription. To determine your Fabric home region, visit Find your Fabric home region.

Asia Pacific Europe Americas Middle East and Africa
Australia East North Europe Brazil South South Africa North
Australia Southeast West Europe Canada Central UAE North
Central Indian France Central Canada East
East Asia Norway East East US
Japan East Switzerland North East US 2
Korea Central Switzerland West North Central US
Southeast Asia UK South South Central US
South India UK West West US
West US 2
West US 3

Consumption rate

Consumption rate for OpenAI language models

Model Deployment Name Context Window (Tokens) Input (Per 1,000 Tokens) Cached Input (Per 1,000 Tokens) Output (Per 1,000 Tokens) Retirement Date
gpt-5-2025-08-07 gpt-5 400,000
Max output: 128,000		 42.02 CU seconds 4.20 CU seconds 336.13 CU seconds TBD
gpt-4.1-2025-04-14 gpt-4.1 128,000
Max output: 32,768		 67.23 CU seconds 16.81 CU seconds 268.91 CU seconds TBD
gpt-4.1-mini-2025-04-14 gpt-4.1-mini 128,000
Max output: 32,768		 13.45 CU seconds 3.36 CU seconds 53.78 CU seconds TBD

Consumption rate for OpenAI embedding models

Models Deployment Name Context (Tokens) Input (Per 1,000 Tokens)
Ada text-embedding-ada-002 8192 3.36 CU seconds

Consumption rate for Text Analytics

Operation Operation Unit of Measure Consumption rate
Language Detection 1,000 text records 33,613.45 CU seconds
Sentiment Analysis 1,000 text records 33,613.45 CU seconds
Key Phrase Extraction 1,000 text records 33,613.45 CU seconds
Personally Identifying Information Entity Recognition 1,000 text records 33,613.45 CU seconds
Named Entity Recognition 1,000 text records 33,613.45 CU seconds
Entity Linking 1,000 text records 33,613.45 CU seconds
Summarization 1,000 text records 67,226.89 CU seconds

Consumption rate for Text Translator

Operation Operation Unit of Measure Consumption rate
Translate 1M Characters 336,134.45 CU seconds
Transliterate 1M Characters 336,134.45 CU seconds

Changes to Foundry Tools in Fabric consumption rate

Consumption rates are subject to change at any time. Microsoft uses reasonable efforts to provide notice via email or through in-product notification. Changes shall be effective on the date stated in the Microsoft Release Notes or the Microsoft Fabric Blog. If any change to an AI service in Fabric Consumption Rate materially increases the Capacity Units (CU) required to use, customers can use the cancellation options available for the chosen payment method.

Monitor the Usage

The workload meter associated with the task determines the charges for prebuilt Foundry Tools in Fabric. For example, if Foundry Tool usage is derived from a Spark workload, the AI usage is grouped together and billed under the Spark billing meter on Fabric Capacity Metrics app.

Note

The billing for prebuilt Foundry Tools does not support the Autoscale Spark billing.

Example

An online shop owner uses SynapseML and Spark to categorize millions of products into relevant categories. Currently, the shop owner applies hard-coded logic to clean and map the raw "product type" to categories. However, the owner plans to switch to use of the new native Fabric OpenAI LLM (Large Language Model) endpoints. This iteratively processes the data against an LLM for each row, and then categorizes the products based on their "product name," "description," "technical details," and so on.

The expected cost for Spark usage is 1000 CUs. The expected cost for OpenAI usage is about 300 CUs.

To test the new logic, first iterate it in a Spark notebook interactive run. For the operation name of the run, use "Notebook Run." The owner expects to see an all-up usage of 1300 CUs under "Notebook Run," with the Spark billing meter accounting for the entire usage.​

Once the shop owner validates the logic, the owner sets up the regular run and expects to see an all-up usage of 1300 CUs under the operation name "Spark Job Scheduled Run," with the Spark billing meter accounting for the entire usage.​

According to Spark compute usage reporting, all Spark related operations are classified as background operations.

Related content

  • Last updated on