
Retrieval-augmented generation (RAG) grounds model output in an organization's own data through a six-step cycle: ingestion and chunking, embedding, indexing, retrieval, augmentation, and generation. On Google Cloud, Gemini Enterprise Agent Platform (Agent Platform) offers three tools for building this cycle: Agent Search, RAG Engine, and Vector Search. For the Professional Machine Learning Engineer exam, choosing among them comes down to how much control an architecture requires and which layer handles indexing, retrieval, and orchestration.
Agent Search is the most managed RAG option on Agent Platform. It supports personalized content recommendations for tailored discovery experiences, and it lets you build a search engine on top of your own data, similar to creating a private search engine over an organization's documents instead of the public web.
To prepare unstructured files for retrieval, Agent Search includes a layout parser that detects structural elements such as text blocks, tables, and headings. Documents do not naturally break into clean pieces for a model to read, so preserving headings and tables as a document is split into chunks keeps that context intact for search and RAG workflows. Agent Search also provides pre-built connectors for Cloud Storage, BigQuery, websites, and Google Drive, allowing you to pull data from where it already lives without building custom ingestion connectors.
Three common use cases appear regularly for Agent Search:
One distinction to keep straight for the Professional Machine Learning Engineer exam is the boundary between Agent Search and Document AI. Although Agent Search detects document elements like tables and headings during chunking, it is designed for semantic search and content retrieval rather than structured field extraction workflows. It finds and ranks relevant passages of text, whereas a dedicated document extraction tool such as Document AI pulls out specific fields like an invoice number.
RAG Engine is a component of Agent Platform built specifically to run the six-step RAG cycle. It operates as a managed orchestration layer that retrieves relevant data and feeds it into the LLM's context, connecting the retrieval and generation steps instead of requiring your team to wire them together manually.
RAG Engine supports a wide range of models, including Gemini and all models available in Model Garden, so an application is not locked into a single model family. It also lets you choose what handles retrieval underneath. While RAG Engine can connect to several backends, the two to know for the exam are Vector Search and Agent Search. You can configure RAG Engine to search embeddings stored in Vector Search, or you can configure it to use Agent Search as its retrieval backend so Agent Search's own search engine performs the lookup. Those two choices are mutually exclusive (one or the other, not both at once), and in either configuration RAG Engine takes the returned results and inserts them into the model as context.
Vector Search is a high-performance vector search engine for storing, indexing, and retrieving embeddings across search and recommendation systems. It organizes embeddings into indexes optimized for similarity search, making it possible to locate the closest matching vectors quickly across large datasets.
Vector Search plays two roles on Agent Platform. First, it serves as the underlying vector database used by Agent Search, performing similarity search beneath Agent Search's managed layer. Second, you can use Vector Search directly when an application requires more control than Agent Search exposes, such as fine-tuning how indexes are built or queried. It can serve as the vector store behind RAG Engine or a custom RAG solution to improve scalability while your team builds the rest of the pipeline around it. Vector Search also supports custom embeddings and specialized retrieval use cases, so you are not restricted to Google Cloud's own embedding models when a workload needs something different.
An application on Google Cloud that requires RAG can follow one of three paths arranged by complexity and control:
When a Professional Machine Learning Engineer scenario calls for managed search over existing Cloud Storage, BigQuery, website, or Google Drive sources, or tailored content recommendations with minimal setup, we would lean toward Agent Search. When the requirement is managed RAG orchestration across Gemini or Model Garden models without wiring retrieval and generation together manually, RAG Engine fits in between. When a team needs custom embedding models, fine-tuned index construction and querying, or a scalable vector store for a custom RAG pipeline, Vector Search is the appropriate building block.
Our Professional Machine Learning Engineer course covers RAG options on Agent Platform alongside Document AI and Model Garden, with practice questions that drill these distinctions.