Rag ADK

Readme.md

Title: Deployment modes in RAG Engine on Gemini Enterprise Agent Platform  |  Google Cloud Documentation

Description: Manage Deployment modes in RAG Engine on Gemini Enterprise Agent Platform. Configure Serverless and Spanner modes to optimize RAG performance and costs.

Source: https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/rag-engine/deployment-modes

Using the UI for RAG Grounding

You can also configure RAG Engine grounding directly through the Google Cloud UI.

1. Model Settings: Select your model (e.g., gemini-3.1-flash-lite) and set your Thinking level and Structured output preferences.
2. Grounding: In the grounding section, you can select tools like Google Search, Google Maps, or RAG Engine (along with Agent Search or Elasticsearch).
3. Region: Make sure your region is set appropriately (e.g., global).
4. Change the Corpus: When using the RAG Engine for grounding in the UI, you can easily change the corpus you are querying against directly from the interface.

---

In Google Cloud, these solve related but different problems:

Option	Best For	Managed AI Workflow	Semantic Search	LLM Integration	Typical Product
RAG	AI chat/Q&A over documents	Yes	Yes	Built-in	Vertex AI RAG Engine
Vector Search	Custom similarity systems	Partial	Yes	You build it	Vertex AI Vector Search
Search	Traditional enterprise/document search	Yes	Limited/Hybrid	Optional	Vertex AI Search

1. Vertex AI RAG Engine

Use when you want:

• Chat with documents
• Gemini grounded on your data
• Minimal infrastructure
• Managed ingestion + retrieval

Google handles:

• chunking
• embeddings
• indexing
• retrieval
• grounding

You mainly do:

rag.retrieval_query(...)

or attach RAG to Gemini tools.

Best for

• Internal AI assistants
• PDF Q&A
• Knowledge base chatbots
• Fastest time-to-market

Pros

• Simplest
• Native Gemini integration
• Least code
• Fully managed

Cons

• Less control
• Less customizable ranking
• Limited advanced retrieval tuning

---

2. Vertex AI Vector Search

Use when you need:

• Custom retrieval systems
• High-scale vector similarity
• Hybrid ranking
• Recommendation engines
• Full control

You manage:

• embeddings
• chunking
• metadata
• indexing pipeline
• retrieval orchestration

Typical flow:

Documents
 → embeddings
 → Vector Search index
 → retrieve neighbors
 → send context to Gemini

Best for

• Large-scale semantic search
• Recommendation systems
• Custom AI architecture
• Multi-modal retrieval
• Production ML infra

Pros

• Maximum flexibility
• High performance
• Enterprise scale
• Custom ranking/filtering

Cons

• More engineering work
• You build RAG orchestration yourself

---

3. Vertex AI Search

Use when you need:

• Google-like enterprise search
• Website search
• Search portal
• Structured + unstructured search

This is closer to:

• enterprise search
• website search
• document discovery

Less focused on LLM orchestration.

Best for

• Search experiences
• Employee portals
• Ecommerce/document search
• Websites

Pros

• Strong search UX
• Search-focused features
• Easy indexing

Cons

• Less flexible for advanced GenAI
• Not optimized primarily for Gemini grounding

---

What most teams should use

Use Vertex AI RAG Engine if:

• You are building AI chat/Q&A
• You want Gemini + documents
• You want fastest implementation
• Your corpus is moderate size

This is probably what you want right now.

---

Use Vector Search if:

• You need custom architecture
• Millions/billions of vectors
• Advanced retrieval tuning
• Hybrid ranking pipelines
• Recommendation/search platform

---

Use Vertex AI Search if:

• You want Google-style enterprise search
• Search UI is primary product
• Not mainly conversational AI

---

Recommended Architecture

Small/Medium AI Assistant

Gemini
 + Vertex AI RAG Engine

Advanced Enterprise AI

Gemini
 + Vertex AI Vector Search
 + Custom retrieval pipeline

Enterprise Search Portal

Vertex AI Search
 + optional Gemini summaries

---

For your current code

Since you're already using:

from vertexai import rag

You are using:

✅ Vertex AI RAG Engine

That is the correct choice if your goal is:

• chatbot over docs
• grounding Gemini
• internal knowledge assistant
• fast implementation

You only need Vector Search directly if you outgrow the managed RAG abstraction.

adk-rag-engine.py

from google import genai
from google.genai import types
import base64
import os

def generate():
  client = genai.Client(
      vertexai=True,
      project="user-xpkurowsfvgo",
      location="us-south1",
  )

  model = "gemini-2.5-flash"
  contents = [
      "What is RAG and why is it helpful?"
  ]
  tools = [
    types.Tool(
      retrieval=types.Retrieval(
        vertex_rag_store=types.VertexRagStore(
          rag_resources=[
            types.VertexRagStoreRagResource(
              rag_corpus="projects/user-xpkurowsfvgo/locations/us-south1/ragCorpora/2305843009213693952"
            )
          ],
          similarity_top_k=20,
        )
      )
    )
  ]

  generate_content_config = types.GenerateContentConfig(
    temperature = 1,
    top_p = 0.95,
    max_output_tokens = 65535,
    safety_settings = [types.SafetySetting(
      category="HARM_CATEGORY_HATE_SPEECH",
      threshold="OFF"
    ),types.SafetySetting(
      category="HARM_CATEGORY_DANGEROUS_CONTENT",
      threshold="OFF"
    ),types.SafetySetting(
      category="HARM_CATEGORY_SEXUALLY_EXPLICIT",
      threshold="OFF"
    ),types.SafetySetting(
      category="HARM_CATEGORY_HARASSMENT",
      threshold="OFF"
    )],
    tools = tools,
  )

  for chunk in client.models.generate_content_stream(
    model = model,
    contents = contents,
    config = generate_content_config,
    ):
    if not chunk.candidates or not chunk.candidates[0].content or not chunk.candidates[0].content.parts:
        continue
    print(chunk.text, end="")

generate()

list-corpora.py

from vertexai import rag

# gcloud services enable cloudresourcemanager.googleapis.com
import vertexai
vertexai.init(project='user-xpkurowsfvgo', location="us-south1")
for corpus in rag.list_corpora():
    print(f"Name: {corpus.name} | Display Name: {corpus.display_name}")