Building an Enterprise RAG Pipeline in Azure with NVIDIA AI Blueprint for RAG and Azure NetApp Files
Transform your enterprise-grade RAG pipeline with NVIDIA AI and Azure NetApp Files. This post highlights the challenges of scaling RAG solutions and introduces NVIDIA's AI Blueprint adapted for Azure. Discover how Azure NetApp Files boosts performance and handles dynamic demands, enabling robust and efficient RAG workloads.Streamlining data discovery for AI/ML with OpenMetadata on AKS and Azure NetApp Files
This article contains a step-by-step guide to deploying OpenMetadata on Azure Kubernetes Service (AKS), using Azure NetApp Files for storage. It also covers the deployment and configuration of PostgreSQL and OpenSearch databases to run externally from the Kubernetes cluster, following OpenMetadata best practices, managed by NetApp® Instaclustr®. This comprehensive tutorial aims to assist Microsoft and NetApp customers in overcoming the challenges of identifying and managing their data for AI/ML purposes. By following this guide, users will achieve a fully functional OpenMetadata instance, enabling efficient data discovery, enhanced collaboration, and robust data governance.
Transform Insurance Industry Workflows Using Generative AI Models and Azure Services
This article highlights an innovative automated solution designed to transform the processing of insurance claim forms for the insurance industry. Previously, underwriters were limited to handling just two to three claims per day, significantly hampering operational efficiency. With the implementation of this solution, companies have achieved a remarkable 60% increase in daily claim processing capacity. Built on Azure services, this architecture revolutionizes the management of claim forms submitted via email by automating critical tasks such as data extraction, classification, summarization, evaluation, and storage. Leveraging the power of AI and machine learning, this solution ensures faster, more accurate claim evaluations, enabling insurance companies to make informed decisions efficiently. The result is enhanced operational scalability, improved customer satisfaction, and a streamlined claims process. Scenario In the insurance industry, claim forms often arrive as email attachments, requiring manual processing to classify, extract, and validate information before it can be stored for analysis and reporting. This solution automates the process by leveraging Azure services to classify, extract, and summarize information from Insurance claim forms. Using Responsible AI evaluation, it ensures the performance of Large Language Models (LLMs) meets high standards. The data is then stored for further analysis and visualization in Power BI, where underwriters can access consumable reports. Architecture Diagram Components Azure Logic Apps: Automates workflows and integrates apps, data, and services. Used here to process emails, extract PDF attachments, and initiate workflows with an Outlook connector for attachment, metadata, and email content extraction. Azure Blob Storage: Stores unstructured data at scale. Used to save insurance claim forms in PDF and metadata/email content in TXT formats. Azure Functions: Serverless compute for event-driven code. Orchestrates workflows across services. Azure Document Intelligence: AI-powered data extraction from documents. Classifies and extracts structured content from ACCORD forms. Azure OpenAI: Provides advanced language models. Summarizes email content for high-level insights. LLM Evaluation Module (Azure AI SDK): Enhances Azure OpenAI summaries by evaluating and refining output quality. Azure AI Foundry: Manages Azure OpenAI deployments and evaluates LLM performance using Responsible AI metrics. Azure Cosmos DB: Globally distributed NoSQL database. Stores JSON outputs from Azure OpenAI and Document Intelligence. Microsoft Power BI: Visualizes Cosmos DB data with interactive reports for underwriters. Workflow Description The workflow for processing claims efficiently leverages a series of Azure services to automate, structure, and analyze data, ensuring a fast, accurate, and scalable claims management system. 1. Email Processing with Azure Logic Apps The process begins with a pre-designed Azure Logic Apps workflow, which automates the intake of PDF claim forms received as email attachments from policyholders. By using prebuilt Outlook connectors, it extracts key details like sender information, email content, metadata, and attachments, organizing the data for smooth claims processing. This automation reduces manual effort, accelerates claim intake, and minimizes data capture errors. 2. Secure Data Storage in Azure Blob Storage Once emails are processed, the necessary PDF attachments, email content, and email metadata are stored securely in Azure Blob Storage. This centralized, scalable repository ensures easy access to raw claim data for subsequent processing. Azure Blob’s structured storage supports efficient file retrieval during later stages, while its scalability can handle growing claim volumes, ensuring data integrity and accessibility throughout the entire claims processing lifecycle. 3. Workflow Orchestration with Azure Functions The entire processing workflow is managed by Azure Functions, which orchestrates serverless tasks such as document classification, data extraction, summarization, and LLM evaluation. This modular architecture enables independent updates and optimizations, ensuring scalability and easier maintenance. Azure Functions streamlines operations, improving the overall efficiency of the claims processing system. a. Document Classification: The next step uses Azure Document Intelligence to classify documents with a custom pretrained model, identifying insurance claim forms. This step ensures the correct extraction methods are applied, reducing misclassification and errors, and eliminating much of the need for manual review. The ability to customize the model also adapts to changes in document formats, ensuring accuracy and efficiency in later processes. b. Content Extraction: Once the insurance form is properly classified, Azure Document Intelligence extracts specific data points from the PDF claim forms, such as claim numbers and policyholder details. The automated extraction process saves time, reduces manual data entry, and improves accuracy, ensuring essential data is available for downstream processing. This capability also helps in organizing the information for efficient claim tracking and report generation. c. Document Intelligence Output Processing: The results are extracted in JSON format and then parsed and organized for storage in Azure Cosmos DB, ensuring that all relevant data is systematically stored for future use. d. Summarizing Content with Azure OpenAI: Once data is extracted, Azure OpenAI generates summaries of email content, highlighting key claim submission details. These summaries make it easier for underwriters and decision-makers to quickly understand the essential points without sifting through extensive raw data. e. Quality Evaluation with LLM Evaluation SDK: After summarization, the quality of the generated content is evaluated using the LLM Evaluation Module in the Azure AI SDK. This evaluation ensures that the content meets accuracy and relevance standards, maintaining high-quality benchmarks and upholding responsible AI practices. Insights from this evaluation guide the refinement and improvement of models used in the workflow. f. LLM Performance Dashboard with Azure AI Foundry: Continuous monitoring of the workflow’s quality metrics is done via the evaluation dashboard in Azure AI Foundry. Key performance indicators like Groundedness, fluency, coherence, and relevance are tracked, ensuring high standards are maintained. This regular monitoring helps quickly identify performance issues and informs model optimizations, supporting the efficiency of the claims processing system. g. Summarization Output Processing: After evaluation, the results from the OpenAI summarization output are parsed and stored in Cosmos DB, ensuring that all relevant data is saved in a structured format for easy access and retrieval. 4. Storing Data in Azure Cosmos DB The structured data, including parsed JSON outputs and summaries, is stored in Azure Cosmos DB, a fully managed, globally distributed NoSQL database. This solution ensures processed claim data is easily accessible for further analysis and reporting. Cosmos DB’s scalability can accommodate increasing claim volumes, while its low-latency access makes it ideal for high-demand environments. Its flexible data model also allows seamless integration with other services and applications, improving the overall efficiency of the claims processing system. 5. Data Visualization with Microsoft Power BI The final step in the workflow involves visualizing the stored data using Microsoft Power BI. This powerful business analytics tool enables underwriters and other stakeholders to create interactive reports and dashboards, providing actionable insights from processed claim data. Power BI’s intuitive interface allows users to explore data in depth, facilitating quick, data-driven decisions. By incorporating Power BI, the insurance company can effectively leverage stored data to drive business outcomes and continuously improve the claims management process. Related Use cases: Healthcare - Patient Intake and Medical Claims Processing: Automating the extraction and processing of patient intake forms and medical claims for faster reimbursement and improved patient care analysis. See the following article for more information on how to implement a solution like this. Financial Services - Loan and Mortgage Application Processing: Streamlining loan application reviews by automatically extracting and summarizing financial data for quicker decision-making. Retail - Supplier Invoice and Purchase Order Processing: Automating invoice and purchase order processing for faster supplier payment approvals and improved financial tracking. Legal contract and Document Review: Automating the classification and extraction of key clauses from legal contracts to enhance compliance and reduce manual review time. See the following article for more information on how to implement a solution like this. Government - Tax Filing and Documentation Processing: Automating the classification and extraction of tax filing data to ensure compliance and improve audit efficiency. To find solution ideas and reference architectures for Azure based solutions curated by Microsoft, go to the Azure Architecture Center and search with keywords like “retail”, “legal”, “healthcare”, etc. You’ll find hundreds of industry-related solutions that can help jumpstart your design process. Contributors: This article is maintained by Microsoft. It was originally written by the following contributors. Principal authors: Manasa Ramalinga| Principal Cloud Solution Architect – US Customer Success Oscar Shimabukuro Kiyan| Senior Cloud Solution Architect – US Customer SuccessAzure AI Foundry, GitHub Copilot, Fabric and more to Analyze usage stats from Utility Invoices
Overview With the introduction of Azure AI Foundry, integrating various AI services to streamline AI solution development and deployment of Agentic AI Workflow solutions like multi-modal, multi-model, dynamic & interactive Agents etc. has become more efficient. The platform offers a range of AI services, including Document Intelligence for extracting data from documents, natural language processing and robust machine learning capabilities, and more. Microsoft Fabric further enhances this ecosystem by providing robust data storage, analytics, and data science tools, enabling seamless data management and analysis. Additionally, Copilot and GitHub Copilot assist developers by offering AI-powered code suggestions and automating repetitive coding tasks, significantly boosting productivity and efficiency. Objectives In this use case, we will use monthly electricity bills from the utilities' website for a year and analyze them using Azure AI services within Azure AI Foundry. The electricity bills is simply an easy start but we could apply it to any other format really. Like say, W-2, I-9, 1099, ISO, EHR etc. By leveraging the Foundry's workflow capabilities, we will streamline the development stages step by step. Initially, we will use Document Intelligence to extract key data such as usage in kilowatts (KW), billed consumption, and other necessary information from each PDF file. This data will then be stored in Microsoft Fabric, where we will utilize its analytics and data science capabilities to process and analyze the information. We will also include a bit of processing steps to include Azure Functions to utilize GitHub Copilot in VS Code. Finally, we will create a Power BI dashboard in Fabric to visually display the analysis, providing insights into electricity usage trends and billing patterns over the year. Utility Invoice sample Building the solution Depicted in the picture are the key Azure and Copilot Services we will use to build the solution. Set up Azure AI Foundry Create a new project in Azure AI Foundry. Add Document Intelligence to your project. You can do this directly within the Foundry portal. Extract documents through Doc Intel Download the PDF files of the power bills and upload them to Azure Blob storage. I used Document Intelligence Studio to create a new project and Train custom models using the files from the Blob storage. Next, in your Azure AI Foundry project, add the Document Intelligence resource by providing the Endpoint URL and Keys. Data Extraction Use Azure Document Intelligence to extract required information from the PDF files. From the resource page in the Doc Intel service in the portal, copy the Endpoint URL and Keys. We will need these to connect the application to the Document Intelligence API. Next, let’s integrate doc intel with the project. In the Azure AI Foundry project, add the Document Intelligence resource by providing the Endpoint URL and Keys. Configure the settings as needed to start using doc intel for extracting data from the PDF documents. We can stay within the Azure AI Foundry portal for most of these steps, but for more advanced configurations, we might need to use the Document Intelligence Studio. GitHub Copilot in VS Code for Azure Functions For processing portions of the output from Doc Intel, what better way to create the Azure Function than in VS Code, especially with the help of GitHub Copilot. Let’s start by installing the Azure Functions extension in VS Code, then create a new function project. GitHub Copilot can assist in writing the code to process the JSON received. Additionally, we can get Copilot to help generate unit tests to ensure the function works correctly. We could use Copilot to explain the code and the tests it generates. Finally, we seamlessly integrate the generated code and unit tests into the Functions app code file, all within VS Code. Notice how we can prompt GitHub Copilot from step 1 of Creating the Workspace to inserting the generated code into the Python file for the Azure Function to testing it and all the way to deploying the Function. Store and Analyze information in Fabric There are many options for storing and analyzing JSON data in Fabric. Lakehouse, Data Warehouse, SQL Database, Power BI Datamart. As our dataset is small, let’s choose either SQL DB or PBI Datamart. PBI Datamart is great for smaller datasets and direct integration with PBI for dashboarding while SQL DB is good for moderate data volumes and supports transactional & analytical workloads. To insert the JSON values derived in the Azure Functions App either called from Logic Apps or directly from the AI Foundry through the API calls into Fabric, let’s explore two approaches. Using REST API and the other Using Functions with Azure SQL DB. Using REST API – Fabric provides APIs that we can call directly from our Function to insert records using HTTP client in the Function’s Python code to send POST requests to the Fabric API endpoints with our JSON data. Using Functions with Azure SQL DB – we can connect it directly from our Function using the SQL client in the Function to execute SQL INSERT statements to add records to the database. While we are at it, we could even get GitHub Copilot to write up the Unit Tests. Here’s a sample: Visualization in Fabric Power BI Let's start with creating visualizations in Fabric using the web version of Power BI for our report, UtilitiesBillAnalysisDashboard. You could use the PBI Desktop version too. Open the PBI Service and navigate to the workspace where you want to create your report. Click on "New" and select "Dataset" to add a new data source. Choose "SQL Server" from the list of data sources and enter "UtilityBillsServer" as the server name and "UtilityBillsDB" as the DB name to establish the connection. Once connected, navigate to the Navigator pane where we can select the table "tblElectricity" and the columns. I’ve shown these in the pictures below. For a clustered column (or bar) chart, let us choose the columns that contain our categorical data (e.g., month, year) and numerical data (e.g., kWh usage, billed amounts). After loading the data into PBI, drag the desired fields into the Values and Axis areas of the clustered column chart visualization. Customize the chart by adjusting the formatting options to enhance readability and insights. We now visualize our data in PBI within Fabric. We may need to do custom sort of the Month column. Let’s do this in the Data view. Select the table and create a new column with the following formula. This will create a custom sort column that we will use as ‘Sum of MonthNumber’ in ascending order. Other visualizations possibilities: Other Possibilities Agents with Custom Copilot Studio Next, you could leverage a custom Copilot to provide personalized energy usage recommendations based on historical data. Start by integrating the Copilot with your existing data pipeline in Azure AI Foundry. The Copilot can analyze electricity consumption patterns stored in your Fabric SQL DB and use ML models to identify optimization opportunities. For instance, it could suggest energy-efficient appliances, optimal usage times, or tips to reduce consumption. These recommendations can be visualized in PBI where users can track progress over time. To implement this, you would need to set up an API endpoint for the Copilot to access the data, train the ML models using Python in VS Code (let GitHub Copilot help you here… you will love it), and deploy the models to Azure using CLI / PowerShell / Bicep / Terraform / ARM or the Azure portal. Finally, connect the Copilot to PBI to visualize the personalized recommendations. Additionally, you could explore using Azure AI Agents for automated anomaly detection and alerts. This agent could monitor electricity bill data for unusual patterns and send notifications when anomalies are detected. Yet another idea would be to implement predictive maintenance for electrical systems, where an AI agent uses predictive analytics to forecast maintenance needs based on the data collected, helping to reduce downtime and improve system reliability. Summary We have built a solution that leveraged the seamless integration of pioneering AI technologies with Microsoft’s end-to-end platform. By leveraging Azure AI Foundry, we have developed a solution that uses Document Intelligence to scan electricity bills, stores the data in Fabric SQL DB, and processes it with Python in Azure Functions in VS Code, assisted by GitHub Copilot. The resulting insights are visualized in Power BI within Fabric. Additionally, we explored potential enhancements using Azure AI Agents and Custom Copilots, showcasing the ease of implementation and the transformative possibilities. Finally, speaking of possibilities – With Gen AI, the only limit is our imagination! Additional resources Explore Azure AI Foundry Start using the Azure AI Foundry SDK Review the Azure AI Foundry documentation and Call Azure Logic Apps as functions using Azure OpenAI Assistants Take the Azure AI Learn courses Learn more about Azure AI Services Document Intelligence: Azure AI Doc Intel GitHub Copilot examples: What can GitHub Copilot do – Examples Explore Microsoft Fabric: Microsoft Fabric Documentation See what you can connect with Azure Logic Apps: Azure Logic Apps Connectors About the Author Pradyumna (Prad) Harish is a Technology leader in the GSI Partner Organization at Microsoft. He has 26 years of experience in Product Engineering, Partner Development, Presales, and Delivery. Responsible for revenue growth through Cloud, AI, Cognitive Services, ML, Data & Analytics, Integration, DevOps, Open Source Software, Enterprise Architecture, IoT, Digital strategies and other innovative areas for business generation and transformation; achieving revenue targets via extensive experience in managing global functions, global accounts, products, and solution architects across over 26 countries.Getting started with the NetApp Connector for Microsoft M365 Copilot and Azure NetApp Files
Imagine a world where your on-premises and enterprise cloud files seamlessly integrate with Microsoft Copilot unleashing AI on your Azure NetApp Files enterprise data, and making your workday smoother and more efficient. Welcome to the future with the NetApp Connector for Microsoft Copilot!Building scalable and persistent AI applications with LangChain, Instaclustr, and Azure NetApp Files
Discover the powerful combination of LangChain and LangGraph for building stateful AI applications and unlock the benefits of using a managed-database service like NetApp® Instaclustr® backed by Azure NetApp Files for seamless data persistence and scalability.High-performance storage for AI Model Training tasks using Azure ML studio with Azure NetApp Files
This article describes how to provide enterprise grade high performance persistent storage with data protection capability for AI Model training tasks using studio compute instances with Azure NetApp Files (ANF).
