Foundation Model Access

Vertex AI gives access to more than a hundred foundation models covering text, code, and other formats, selectable per task rather than locked to one model. A Beam agent passes a prompt to whichever model an approved rule assigns for that task, such as a specific model for drafting replies and another for summarizing documents, and reads back the model's response. It does not choose or swap models on its own. A response that comes back empty, errors out, or scores low on the model's own confidence measure is routed to a person to handle directly.

Conversational Search

Vertex AI's Search and Conversation tools let an application answer a question by pulling from a connected set of documents rather than the model's general training alone. A Beam agent forwards a customer or internal question to an approved conversational search configuration and returns the grounded answer it gets back, along with the source document it cites. It does not maintain the underlying document set. A question the search cannot find a source for, or an answer citing a document marked outdated, is routed to a person to answer directly.

Prediction Endpoints

A model deployed on Vertex AI is reachable through a prediction endpoint, which takes an input and returns a scored output, such as a classification or a ranked list. A Beam agent calls the endpoint an approved rule points to for a given task, reads the returned score, and acts on it, such as tagging a record above an agreed confidence level. It does not deploy or retrain the model behind the endpoint. Predictions below the confidence level, or an endpoint returning an error, are routed to a person to decide the case manually.