Azure OpenAI On Your Data Feature
Jump to navigation
Jump to search
A Azure OpenAI On Your Data Feature is a RAG service for Azure OpenAI Services (that allows developers to connect, ingest, and ground their enterprise data to create personalized AI assistants).
- Context:
- It can (typically) run advanced AI models like GPT-35-Turbo and GPT-4 on an organization's own enterprise data without needing to train or fine-tune the models.
- It can (often) be accessed via REST API, SDK, and a web-based interface in Azure OpenAI Studio.
- It can allow users to chat and analyze their data with greater accuracy, specifying sources to support responses based on the latest information in designated data sources.
- It can range from being a solution for creating a web app connected to data for enhanced chat solutions to deploying it directly as a copilot in the Copilot Studio (preview).
- It can support various data formats, including .txt, .md, .html, .docx, .pptx, and .pdf files.
- It can integrate data sources like Azure AI Search, Vector Database in Azure Cosmos DB for MongoDB, Azure Blob Storage (preview), uploaded files (preview), URL/web address (preview), and Elasticsearch (preview).
- It can offer different search types, including keyword search, semantic search, vector search (using Ada embedding models in selected regions), and hybrid search options.
- It can enhance user comprehension, expedite task completion, improve operational efficiency, and aid in decision-making.
- It can include security features like document-level access control and integration with Azure Role-Based Access Control (RBAC).
- It can customize various parameters like chunk size, number of retrieved documents, and response strictness to optimize performance for specific use cases.
- ...
- Example(s):
- Counter-Example(s):
- ...
- See: Azure AI Search, GPT-4, Azure Role-Based Access Control (RBAC).
References
2024
- (Microsoft, 2024]]) => "Azure OpenAI On Your Data."
- NOTES:
- **Overview and Purpose**: Azure OpenAI On Your Data enables developers to connect, ingest, and ground enterprise data to create personalized AI copilots rapidly. It aims to enhance comprehension, expedite tasks, improve operational efficiency, and aid decision-making without requiring model training or fine-tuning.
- **Supported AI Models**: The service allows running advanced AI models like GPT-35-Turbo and GPT-4 on enterprise data, enabling accurate data analysis and chat functionalities using a REST API, SDK, or Azure OpenAI Studio.
- NOTES:
3. **Data Integration**:
- It supports various data formats, including .txt, .md, .html, .docx, .pptx, and .pdf. The service uses Azure AI Search for indexing and managing data from local machines, URLs, and Azure Blob Storage.
4. **Role-Based Access Control (RBAC)**:
- Azure RBAC ensures secure data integration by requiring specific roles for data source addition, supporting document-level access control with security filters based on user Microsoft Entra group membership.
5. **Search Types and Options**:
- The service offers keyword search, semantic search, and vector search using Ada embedding models. Hybrid search options combine these methods, enhancing retrieval accuracy and relevance.
6. **Intelligent Search and Indexing**:
- Azure OpenAI On Your Data features intelligent search with default semantic and keyword search. Index field mapping and document chunking optimize data ingestion and response accuracy, with customizable parameters for chunk size, strictness, and retrieved documents.
7. **Deployment and Integration**:
- Solutions can be deployed as a copilot, Teams app, or web app directly from Azure OpenAI Studio. It supports streaming data, enabling real-time interactions and enhancing user experience.
8. **Best Practices and Troubleshooting**:
- Recommendations include adjusting chunk size and runtime parameters for optimal performance, clearing conversation history for unrelated queries, and employing prompt engineering techniques. Troubleshooting includes addressing ingestion, quota, and permissions issues for smooth operation.