A Generative AI Engineer is designing a RAG application for answering user questions on technical regulations as they learn a new sport. What are the steps needed to build this RAG application and deploy it?

The correct sequence for building and deploying a Retrieval-Augmented Generation (RAG) application follows a logical progression from data preparation to deployment. First, the knowledge base (documents) is ingested and indexed into a vector database like Databricks Vector Search. This is the offline preparation stage. The online, or inference, stage begins when a user submits a query. The system retrieves relevant documents from the vector store and passes them, along with the original query, to an LLM, which then generates a response. Following generation, the entire application is evaluated for quality and accuracy. Once it meets performance criteria, it is deployed for end-user access, typically using a tool like Databricks Model Serving.

A. This option is incorrect because the evaluation step must occur after the LLM generates a response, as the quality of the final output is a key metric.

C. This option is critically incomplete. It omits the essential online RAG steps of handling user queries, retrieving documents, and generating responses with an LLM.

D. This option incorrectly places user queries before the foundational steps of document ingestion and indexing. The knowledge base must be built before it can be queried.

1. Databricks Documentation, "What is Retrieval-Augmented Generation (RAG)?": This document outlines the RAG architecture, showing the initial offline step of indexing data into a vector database, followed by the online steps of retrieving documents based on a user query and then sending the augmented prompt to an LLM to generate an answer. This supports the sequence: Ingest -> Index -> Query -> Retrieve -> Generate.

Reference: Databricks AI and Machine Learning Guide > Generative AI > RAG > What is Retrieval-Augmented Generation (RAG)?, Architecture section.

2. Databricks Documentation, "LLMOps: LLM development lifecycle": This guide describes the end-to-end lifecycle for LLM applications. It places evaluation after the development and prompt engineering phase (which includes generation) and before the deployment phase. This confirms the sequence: ...Generate -> Evaluate -> Deploy.

Reference: Databricks AI and Machine Learning Guide > LLMOps > LLM development lifecycle, "Step 4. Evaluate and refine the application" and "Step 5. Deploy the LLM".

3. Databricks Blog, "Building RAG Applications with the Databricks Platform": This official blog post details the RAG workflow, starting with creating a knowledge base (ingest/index), followed by the retrieval and generation chain, and then discussing the importance of evaluation before deployment using Model Serving.

Reference: Databricks Blog, "Building RAG Applications with the Databricks Platform", sections "Creating a knowledge base with Databricks Vector Search" and "Deploying the RAG Chain to a scalable real-time endpoint".