
Finding the items most similar to a query across millions of records in a few hundred milliseconds is the problem that embeddings and a vector database are built to solve. On Google Cloud, Agent Retrieval, formerly Vector Search 2.0, handles this workload by storing items as data objects in a collection and serving approximate nearest neighbor (ANN) queries. For the Professional Agentic Architect exam, the key architectural distinction is knowing when a workload requires similarity search across a large collection and when the inputs are already known, in which case a single multimodal Gemini request is the appropriate choice.
Similarity search over a large archive depends on separating upfront indexing from query-time lookup. Each record in the collection is converted into an embedding once, and that embedding is indexed. When a new query arrives, the system compares a single query embedding against the index using approximate nearest neighbor (ANN) search rather than inspecting every stored record individually.
With Agent Retrieval, you store items as data objects in a collection, and the service can generate their embeddings automatically before serving ANN queries against the index. Because the collection is embedded once and each query compares a single embedding against the index, this design is what makes low-latency similarity search over millions of items practical.
Several familiar patterns break down as a collection grows:
Embeddings address both sides of the problem. The expensive work of embedding the collection happens once, ahead of time, and the comparison at query time is based on meaning rather than shared words.
Not every task that involves grounding an agent in stored files is a search problem. Similarity search is for discovering which items in a large collection match a query. When the specific inputs are already known, there is nothing to search for, and the work becomes a direct comparison rather than retrieval. Agent Retrieval adds nothing in that situation, because its job is finding items, not examining ones already identified.
Gemini accepts images and text in the same request, so a single call can reason across both kinds of input together. Files stored in Cloud Storage can be passed by their Cloud Storage URIs, which avoids inline size limits for large image files.
A model can only check what it actually receives. Any step that replaces an original input with a summary, a description, or an extraction leaves the model unable to verify details that the replacement left out, and splitting inputs across separate models means no single model sees all of them. Training or fine-tuning a custom model is also a much heavier path, adding training data, upkeep, and cost that a general multimodal model does not require.
When designing a grounding architecture for the Professional Agentic Architect exam, we start by checking whether the agent must discover unknown matches in a large collection or work directly with inputs that are already identified. Discovery across a large collection is the job of embeddings and ANN queries in Agent Retrieval. Work with known inputs usually goes best when the model receives those inputs themselves, in their original form.
Our Professional Agentic Architect practice exams test embeddings and vector search with Agent Retrieval alongside multimodal Gemini requests and grounding agents in data, with detailed explanations that drill these distinctions.