Merge pull request #5109 from MicrosoftDocs/main

5/21/2025 AM Publish

Author: Taojunshen

articles/ai-foundry/how-to/flow-develop.md

@@ -105,7 +105,7 @@ In addition to inline editing the node in the default view, you can also turn on
 
 ### Flow input and output
 
-Flow input is the data passed into the flow as a whole. Define the input schema by specifying the name and type.  Set the input value of each input to test the flow. You can reference the flow input later in the flow nodes using `${input.[input name]}` syntax. 
+Flow input is the data passed into the flow as a whole. Define the input schema by specifying the name and type.  Set the input value of each input to test the flow. You can reference the flow input later in the flow nodes using `${inputs.[input name]}` syntax. 
 
 Flow output is the data produced by the flow as a whole, which summarizes the results of the flow execution. You can view and export the output table after the flow run or batch run is completed.  Define flow output value by referencing the flow single node output using syntax `${[node name].output}` or `${[node name].output.[field name]}`.
 
@@ -120,7 +120,7 @@ Prompt Flow offers not just a streamlined way to execute the flow, but it also b
 At its core, conditional control provides the capability to associate each node in a flow with an **activate config**. This configuration is essentially a "when" statement that determines when a node should be executed. The power of this feature is realized when you have complex flows where the execution of certain tasks depends on the outcome of previous tasks. By using the conditional control, you can configure your specific nodes to execute only when the specified conditions are met.
 
 Specifically, you can set the activate config for a node by selecting the **Activate config** button in the node card. You can add "when" statement and set the condition.
-You can set the conditions by referencing the flow input, or node output. For example, you can set the condition `${input.[input name]}` as specific value or `${[node name].output}` as specific value. 
+You can set the conditions by referencing the flow input, or node output. For example, you can set the condition `${inputs.[input name]}` as specific value or `${[node name].output}` as specific value. 
 
 If the condition isn't met, the node is skipped. The node status is shown as "Bypassed".
 

articles/ai-services/agents/overview.md

@@ -16,7 +16,7 @@ ms.custom: azure-ai-agents
 
 Most businesses don’t want just chatbots - they want automation that's faster and with fewer errors. That might mean summarizing documents, processing invoices, managing support tickets, or publishing blog posts. In all cases, the goal is the same: freeing people and resources to focus on higher-value work by offloading repetitive and predictable tasks.
 
-Large language models (LLMs) opened the door to a new type of automation with systems that can understand unstructured data, make decisions, and generate content. In practice, it can difficult for businesses to move beyond demos and into production. LLMs can drift, be incorrect, and lack accountability. Without visibility, policy enforcement, and orchestration, these models are difficult to trust in real business workflows.
+Large language models (LLMs) opened the door to a new type of automation with systems that can understand unstructured data, make decisions, and generate content. In practice, it can be difficult for businesses to move beyond demos and into production. LLMs can drift, be incorrect, and lack accountability. Without visibility, policy enforcement, and orchestration, these models are difficult to trust in real business workflows.
 
 :::row:::
     :::column span="1":::
@@ -119,4 +119,4 @@ Start with the [environment setup](./environment-setup.md) and [quickstart](./qu
 
 ## Next steps
 
-Learn more about the [models that power agents](./concepts/model-region-support.md).
+Learn more about the [models that power agents](./concepts/model-region-support.md).

articles/ai-services/agents/quotas-limits.md

@@ -26,6 +26,8 @@ The following sections provide you with a guide to the default quotas and limits
 | Max size for all uploaded files for agents |100 GB |  
 | agents token limit | 2,000,000 token limit |
 
+The 2,000,000 agent limit refers to the maximum number of distinct Agent resources that can be created within a single Azure subscription per region. It does not apply to threads or token usage.
+
 ## Quotas and limits for Azure OpenAI models
 
 See the [Azure OpenAI](../openai/quotas-limits.md) for current quotas and limits for the Azure OpenAI models that you can use with Azure AI Foundry Agent Service. 

articles/ai-services/document-intelligence/service-limits.md

@@ -144,9 +144,9 @@ Document Intelligence billing is calculated monthly based on the model type and
 | Adjustable | No | No |
 | **Custom neural model train** | 10 hours per month <sup>5</sup> | no limit (pay by the hour), start with 10 free hours each month |
 | Adjustable | No |Yes <sup>3</sup>|
-| **Max number of pages (Training) * Classifier** | 10,000 | 10,000 (default value) |
+| **Max number of pages (Training) * Classifier** | 25,000 | 25,000 (default value) |
 | Adjustable | No | No |
-| **Max number of document types (classes) * Classifier** | 500 | 500 (default value) |
+| **Max number of document types (classes) * Classifier** | 1000 | 1000 (default value) |
 | Adjustable | No | No |
 | **Training dataset size * Classifier** | 1GB | 2GB (default value) |
 | Adjustable | No | No |

articles/ai-services/openai/how-to/realtime-audio-websockets.md

@@ -60,7 +60,7 @@ You can construct a full request URI by concatenating:
 The following example is a well-constructed `/realtime` request URI:
 
 ```http
-wss://my-eastus2-openai-resource.openai.azure.com/openai/realtime?api-version=2024-12-17&deployment=gpt-4o-mini-realtime-preview-deployment-name
+wss://my-eastus2-openai-resource.openai.azure.com/openai/realtime?api-version=2025-04-01-preview&deployment=gpt-4o-mini-realtime-preview-deployment-name
 ```
 
 To authenticate:

articles/ai-services/openai/how-to/reinforcement-fine-tuning.md

@@ -1,5 +1,5 @@
 ---
-title: 'Customize o4-mini model with Azure OpenAI and reinforcement fine-tuning'
+title: 'Customize o4-mini model with Azure OpenAI and reinforcement fine-tuning (Preview)'
 description: Learn how to use reinforcement fine-tuning with Azure OpenAI
 manager: nitinme
 ms.service: azure-ai-openai
@@ -10,7 +10,7 @@ author: mrbullwinkle
 ms.author: mbullwin
 ---
 
-# Reinforcement fine-tuning (RFT) with Azure OpenAI o4-mini
+# Reinforcement fine-tuning (RFT) with Azure OpenAI o4-mini (Preview)
 
 Reinforcement fine-tuning (RFT) is a technique for improving reasoning models like o4-mini by training them through a reward-based process, rather than relying only on labeled data. By using feedback or "rewards" to guide learning, RFT helps models develop better reasoning and problem-solving skills, especially in cases where labeled examples are limited or complex behaviors are desired.
 
@@ -404,4 +404,4 @@ We also provide a grader check API that you can use to check the validity of you
 
 Aim for a few hundred examples initially and consider scaling up to around 1,000 examples if necessary. The dataset should be balanced, in terms of classes predicted, to avoid bias and ensure generalization.
 
-For the prompts, make sure to provide clear and detailed instructions, including specifying the response format and any constraints on the outputs (e.g. minimum length for explanations, only respond with true/false etc.)
+For the prompts, make sure to provide clear and detailed instructions, including specifying the response format and any constraints on the outputs (e.g. minimum length for explanations, only respond with true/false etc.)

articles/ai-services/openai/includes/fine-tune-models.md

@@ -25,6 +25,7 @@ manager: nitinme
 | `gpt-4.1` (2025-04-14) | North Central US <br> Sweden Central | ✅ | Input: 128,000 <br> Output: 16,384 <br> Training example context length: 65,536 | May 2024 | Text & Vision to Text |
 | `gpt-4.1-mini` (2025-04-14) | North Central US <br> Sweden Central | ✅ | Input: 128,000 <br> Output: 16,384 <br> Training example context length: 65,536 | May 2024 | Text to Text |
 | `gpt-4.1-nano` (2025-04-14) | North Central US <br> Sweden Central | ✅ | Input: 128,000 <br> Output: 16,384 <br> Training example context length: 32,768 | May 2024 | Text to Text |
+| `o4-mini` (2025-04-16) | East US2 <br> Sweden Central | - | Input: 128,000 <br> Output: 16,384 <br> Training example context length: 65,536 | May 2024 | Text to Text |
 
 > [!NOTE]
 > **Global** training (in Public Preview) provides [more affordable](https://aka.ms/aoai-pricing) training per-token, but does not offer [data residency](https://aka.ms/data-residency). It is currently available to Azure OpenAI resources in the following regions, with more regions coming soon:

articles/ai-services/openai/includes/fine-tuning-studio.md

@@ -115,7 +115,7 @@ The first step is to confirm you model choice and the training method. Not all m
 
 - **Supervised Fine Tuning** (SFT): supported by all non-reasoning models.
 - **Direct Preference Optimization (Preview)** ([DPO](../how-to/fine-tuning-direct-preference-optimization.md)): supported by GPT-4o.
-- **Reinforcement Fine Tuning (Preview)** (RFT): supported by reasoning models, like o4-mini.
+- **Reinforcement Fine Tuning (Preview)** ([RFT](../how-to/reinforcement-fine-tuning.md)): supported by reasoning models, like o4-mini.
 
 When selecting the model, you may also select a [previously fine-tuned model](#continuous-fine-tuning).
 

articles/ai-services/openai/realtime-audio-quickstart.md

@@ -62,3 +62,4 @@ Support for the Realtime API was first added in API version `2024-10-01-preview`
 * Learn more about [How to use the Realtime API](./how-to/realtime-audio.md)
 * See the [Realtime API reference](./realtime-audio-reference.md)
 * Learn more about Azure OpenAI [quotas and limits](quotas-limits.md)
+* Learn more about [Language and voice support for the Speech service](../../ai-services/speech-service/language-support.md)