Learn to Build Your Own AI Apps with Azure Cosmos DB! Part 2

Danny, the event planner, opens the session by asking participants to review the code of conduct for a respectful environment. He introduces Jasine Greenway, who will guide the audience through part two of the Azure Cosmos DB developer guide. The session is recorded and will be available on the Microsoft Reactor YouTube channel. Jasine discusses the importance of speed, data handling, cloud solutions, natural language interaction, accuracy, scalability, security, and cost-effectiveness for AI applications. She recaps the progress made in the previous session, where they worked on the front end and experimented with Azure AI Open API endpoints.

Jasine explains the concept of embeddings in machine learning, which is a method to convert text, images, or any data into a form that can be input into a model. She discusses how embeddings can be used to create a searchable product base through chat. The session covers the use of embeddings to understand the relationship between different data points and how Azure Cosmos DB stores these vectors for the application. Jasine also touches on the idea of using natural language to interact with databases and the importance of providing meaningful and fast responses to users.

The presentation moves on to the existing notebook labs, where the participants are reminded that they have access to Azure OpenAI through a proxy for the hackathon. The lab involves setting up an endpoint, creating clients, connecting to Azure OpenAI, and creating vector representations of products stored in Cosmos DB. Jasine demonstrates how to use the embeddings deployment to vectorize and store embeddings in documents. She also differentiates between the GPT model used for chat completions and the model used for creating embeddings, which is text embedding Ada 002.

Jasine guides the audience through creating a vector index, which is essential for vector search in Azure Cosmos DB. She uses the term 'IVF' (Inverted File) to describe the type of index being used. The backend API is then explored, with a focus on how to run it locally and the necessary steps for deployment. The session also covers how to use the Azure Cosmos DB SDK to perform vector searches and how to use the completion model GPT-3.5 for tailored responses to user queries.

The session continues with a focus on Lang Chain, an LLM (Large Language Model) orchestrator that simplifies the process of working with multiple AI packages and libraries. Jasine demonstrates how to use Lang Chain to create clients for completion and embeddings, and how to format documents into JSON for API use. She also discusses the process of running the backend API locally and the steps for deploying it in a container.

Jasine explains the benefits of deploying a backend application as a container, emphasizing consistency and the ease of deployment across different environments. She walks through the Docker steps, showing how to install requirements and serve the API package. The session also covers how to run the application locally and the commands needed to deploy it to Azure Container Apps in the cloud.

The presentation concludes with the final deployment of the application using Bicep, a deployment tool. Jasine encounters some issues with her application and invites participants to share their experiences and solutions. She also mentions an additional demo, which seems to be a local application with a chat interface, and discusses the use of embeddings and completion models in the application. The session ends with a Q&A, where Jasine answers questions about vector embeddings versus completion models and provides a link for the ongoing hackathon.

Jasine addresses the final questions from the audience, clarifying the difference between vector embeddings and completion models. She emphasizes that completion models are for text generation, while vector embeddings are for creating similarity between data points. Jasine also provides information about the hackathon and encourages participants to join the Discord channel for further interaction and support. She wraps up the session by thanking everyone for their time and patience and looks forward to the next interaction on Discord.