feat(google): add VertexRAGRetrieval provider tool

[youpesh]

Summary

Adds a VertexRAGRetrieval provider tool that enables server-side retrieval via Vertex AI RAG Engine during Gemini inference. Retrieval happens in a single inference pass with no tool-call round-trip — critical for low-latency voice agents using native audio.

Usage

from livekit.plugins import google

agent = MyAgent(
    llm=google.realtime.RealtimeModel(
        model="gemini-live-2.5-flash-native-audio",
        vertexai=True,
    ),
    tools=[
        google.tools.VertexRAGRetrieval(
            rag_resources=["projects/123/locations/us-central1/ragCorpora/456"],
            similarity_top_k=3,
        ),
    ],
)

Motivation

Vertex AI RAG Engine works like Google Search grounding — retrieval is handled server-side during inference. With function-calling RAG, Gemini native audio requires 2 inference passes (~3-5s). With server-side retrieval, it's a single pass (~0.5s).

The Retrieval / VertexRagStore types already exist in google.genai.types and the Gemini Live API supports them, but there was no provider tool wrapper to use them through the LiveKit agents framework.

Changes

• Added VertexRAGRetrieval dataclass to tools.py following the same GeminiTool pattern as GoogleSearch, FileSearch, etc.

Closes #5221

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 2 additional findings.

[tinalenguyen]

LGTM, thank you for the PR!!