The correct option is The editor content is tokenized then sent to a service proxy that applies filters then the tokens are forwarded to a hosted LLM and the completion is returned to the IDE.

This reflects the standard flow for AI coding assistants. The editor collects the relevant context and converts it into tokens that the model can process. A proxy applies policy and safety filtering on the way in and on the way out and it also manages privacy controls and routing. The large language model runs as a hosted service and returns a completion that the plugin renders in the IDE.

Only request metadata is transmitted and the assistant never sends source code or tokens is incorrect because the model needs code context that is encoded as tokens to generate a meaningful suggestion. Telemetry can be limited but some tokenized content must be sent for inference.

The IDE publishes the request to Pub/Sub then a Cloud Function compiles the code and writes results to BigQuery which the plugin reads as the completion is incorrect because this describes a data pipeline that is not suited for low latency inference and it does not involve calling a hosted language model to produce a completion.

The language model is bundled inside the IDE so no network requests are made during completion is incorrect because modern general purpose models are typically hosted due to their size and compute needs and most IDE assistants rely on network calls to managed endpoints.

Full AWS Practitioner Certification Question