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

Danny, the event planner for the New York Reactor Space, welcomes everyone and introduces Jasine Greenway, the speaker for the session on Azure Cosmos DB developer guide. The session includes a review of the code of conduct, an overview of the session's structure, and a mention of the ongoing hackathon. Jasine discusses the modern requirements for AI applications, such as speed, data handling, cloud solutions, natural language interaction, accuracy, scalability, security, and cost-effectiveness.

The presentation moves on to discuss the backend development of the web application, which was deployed to Azure App Service. The focus is on using Azure OpenAI service for creating embeddings, which are a way to convert text into a format that can be processed by machine learning models. The concept of vectors and embedding is introduced as a means to build a searchable product base through chat, emphasizing the importance of vector representation for data like images, audio, and text.

Jasine provides a deeper look into the existing notebook labs, focusing on how to use embeddings to create a product base for a searchable database. The process involves setting up an endpoint, creating clients, and connecting to Azure OpenAI. The use of different models for chat completions and embeddings is explained, with a focus on the Text Embedding Ada model for creating vectors. The concept of vectorizing and storing embeddings in documents is also covered.

The session continues with creating a vector index, which is essential for performing vector searches in Azure Cosmos DB. Different types of indexes are discussed, with IVF being the default choice. A demonstration of searching for products using vector representations is shown, highlighting how similar items can be found based on vector similarity. The concept of retrieval augmented generation is introduced, which combines vector databases with completion models to create tailored responses.

Jasine opens up lab five to work with the backend API, showcasing the use of Lang Chain, an LLM (Large Language Model) orchestrator, to manage different AI packages and libraries. The process of setting up the backend API locally, creating clients, and connecting to Azure OpenAI is detailed. The use of Lang Chain to format documents into JSON and the importance of setting up the Python environment for development are also discussed.

The paragraph outlines the steps for deploying the backend API in a container using Docker. It emphasizes the benefits of containerization, such as consistency and ease of deployment across different environments. The process includes installing requirements, using a Dockerfile to manage containers, and deploying the application to Azure Container Apps. The use of Azure Cloud Shell for executing commands is also mentioned.

Jasine clarifies the difference between vector embeddings and interacting with completion models. Completion models are used for text generation, such as creating resumes, while vector embeddings are for creating similarity between data points, which is useful for searches and is typically stored in a vector database. The application of these concepts in a chat application is discussed, where user input is turned into a vector for database comparison.

The session concludes with a Q&A segment where Jasine addresses questions about the use of Docker Desktop for local container testing and the difference between vector embeddings and completion models. She also provides information on how to join the hackathon and Discord for further interaction and support. The importance of community engagement and the opportunity to learn and compete in the hackathon are highlighted before signing off.