A logistics company wants to use a generative AI (gen AI) agent to automatically check real-time inventory levels across its warehouses and adjust delivery schedules. The gen AI agent needs access to internal inventory dat a. They want the most cost-effective solution. What should the organization do?

The most effective and scalable solution is to build a generative AI agent using Vertex AI that is equipped with tools to interact with live data. The internal inventory data can be stored in a suitable Google Cloud database (e.g., Cloud SQL, Spanner). The Vertex AI agent can then be granted a tool that allows it to query this database in real-time to fetch current inventory levels and another tool to adjust schedules. This architectural pattern, often referred to as tool use or a form of Retrieval-Augmented Generation (RAG), is cost-effective because it leverages the model's reasoning capabilities without requiring constant, expensive retraining or fine-tuning to keep up with rapidly changing data.

A. Building a custom API is a component of a solution, not a replacement for the gen AI agent, which is a core requirement of the company.

B. Pre-built gen AI chatbots are generally not designed to securely access and act upon private, real-time internal enterprise data without significant, costly customization.

C. Fine-tuning is unsuitable for real-time data. It would make the model's knowledge static based on the sample data, which would become outdated instantly.

1. Google Cloud Documentation

"Overview of Vertex AI Agents": "Vertex AI Agents lets developers build and deploy enterprise-grade generative AI-powered applications... These applications

called agents

are based on foundation models and are enhanced with developer-provided tools that give agents the ability to interact with external services." This directly supports the architecture of using an agent with tools to access external data sources like a database.

2. Google Cloud Documentation

"Introduction to grounding": "Grounding is the process of connecting model output to verifiable sources of information... Grounding helps models produce more accurate and relevant responses by accessing and processing information that wasn't included in their training data." This principle underlies option D

where the agent grounds its responses and actions in the live inventory database.

3. Google Cloud Blog

"A guide to LLM fine-tuning vs. RAG for enterprise generative AI": This resource explains the trade-offs. For scenarios requiring access to rapidly changing

real-time information (like inventory)

RAG (the pattern described in option D) is presented as the superior and more cost-effective approach compared to fine-tuning (option C). It states

"RAG is ideal for applications that require up-to-date information or access to a large

dynamic knowledge base."