Making a request

Anthropic Vertex SDK Study Notes

**I. Setup and Initialization**

• Installation: Install the Anthropic Python SDK using the magic command: %pip install anthropic-vertex.
• Client Creation: Import AnthropicVertex and create an instance by passing two required keyword arguments:
• region: Specifies the geographical region (e. g. , "global").
• project_id: Your specific Google Cloud project ID.
• Model Definition: Define and assign the specific model version to a variable (e. g. , claude-sonnet-4@20250514).

**II. The client. create Function**

The core function for making requests requires three mandatory keyword arguments:

• model: The name of the model to run (the variable defined in the setup).
• max_tokens: Sets a maximum budget for the generated response.
• Key Concept: This acts as a safety mechanism, not a target. The model will generate whatever response it deems appropriate, but generation will automatically stop if the limit is reached.
• messages: A list representing the exchange between users and the model.

**III. Message Types**

Messages are structured exchanges, similar to a chat application, and fall into two roles:

• User Message: Contains text authored by a person (the user or a developer).
• Assistant Message: Contains text that has been produced by the model and sent back to the user.

**IV. Making a Request**

• Structure: The messages argument must be a list of input messages.
• Single Message Format: A single message is a dictionary containing:
• role: The type of message (e. g. , "user").
• content: The actual string/text to be sent to the model.
• Accessing Output: The response object is deeply nested. To extract only the generated text, access the path: message. content[0]. text.

Takeaways

• Initialization requires creating an AnthropicVertex client instance by providing the required region and project_id.
• The core request function (client. create) mandates three arguments: the model name, the list of messages, and max_tokens.
• max_tokens serves as a safety mechanism, automatically stopping generation when the limit is reached, rather than acting as a target.
• Messages must be structured as a list of dictionaries, each defining a role (user or assistant) and the corresponding content.
• To retrieve the final generated text from the response object, access the deeply nested path: message. content[0]. text.