A Generative AI Engineer is creating an LLM-powered application that will need access to up-to-date news articles and stock prices. The design…

The scenario requires the LLM system to perform actions on external, real-time data sources: querying a structured database (Delta tables) and searching the unstructured web. An agent-based architecture is the most suitable design pattern for this. An LLM agent uses a reasoning engine to determine which "tools" to use to fulfill a request. In this case, the engineer would create a tool for executing SQL queries against the Delta tables and another tool for performing web searches. The agent would then orchestrate these tools, retrieve the necessary real-time data, and provide it as context to the LLM to generate a final, informed response.

A. This describes a specific, linear workflow, not a flexible system architecture. It fails to explain how the LLM initially accesses the "latest news articles" from the web.

B. This is another example of a specific use case (investigating volatility) rather than a robust, general-purpose architecture that can handle various queries about news and stocks.

C. A RAG architecture with a vector store is ill-suited for highly dynamic, real-time data like stock prices and breaking news due to the latency and cost of continuous ingestion and indexing.

1. Databricks Documentation, "LLM agent on Databricks": This documentation explains the concept of LLM agents, stating, "An agent has access to a set of tools and can reason through which tools to use in which order to answer a question." It explicitly mentions tools for looking up data in a database or accessing a web API, which directly maps to the question's requirements.

Source: Databricks Official Documentation > Machine Learning > Generative AI and LLMs > LLM agent on Databricks.

2. Databricks E-book, "The Big Book of MLOps": In the chapter on LLMOps, the book distinguishes between different LLM application patterns. It describes agentic workflows as systems that "can take actions, such as querying a database, calling an API, or running a piece of code." This aligns perfectly with the need to query Delta tables and search the web.

Source: Databricks E-books, "The Big Book of MLOps", Chapter 6: LLMOps, Section: "Common LLM Application Patterns".

3. Databricks Blog, "Building a Q&A Bot on Your Own Data With Llama 2 & Databricks": While focused on RAG, this article implicitly highlights its limitations for real-time data. It states RAG is for "asking questions of your own private data," which is typically a static or semi-static corpus. This contrasts with the agent pattern needed for live, external data sources like web search and real-time databases.

Source: Databricks Official Blog, August 2, 2023, "Building a Q&A Bot on Your Own Data With Llama 2 & Databricks".