gemini_response = model.generate_content(prompt)
print(gemini_response.text)  # (1)!
```

1. Vertex AI Search provides retrieval, Gemini generates answer

```python
# Safety settings (content filtering)
from vertexai.generative_models import HarmCategory, HarmBlockThreshold

safety_settings = {
    HarmCategory.HARM_CATEGORY_HATE_SPEECH: HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,
    HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: HarmBlockThreshold.BLOCK_MEDIUM_AND_ABOVE,
    HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: HarmBlockThreshold.BLOCK_LOW_AND_ABOVE,
    HarmCategory.HARM_CATEGORY_HARASSMENT: HarmBlockThreshold.BLOCK_LOW_AND_ABOVE
}  # (1)!

model = GenerativeModel("gemini-2.0-flash-exp")

response = model.generate_content(
    "User input here",
    safety_settings=safety_settings
)

# Check if blocked
if response.candidates[0].finish_reason == "SAFETY":
    print("Content blocked by safety filters")  # (2)!
```

1. Thresholds: BLOCK_NONE, BLOCK_LOW_AND_ABOVE, BLOCK_MEDIUM_AND_ABOVE, BLOCK_ONLY_HIGH
2. Safety ratings in response metadata

```python
# Custom model deployment (AutoML)
from google.cloud import aiplatform

aiplatform.init(project="my-project", location="us-central1")

# Deploy custom model
model = aiplatform.Model.upload(
    display_name="my-custom-model",
    artifact_uri="gs://my-bucket/model/",
    serving_container_image_uri="us-docker.pkg.dev/vertex-ai/prediction/tf2-cpu.2-13:latest"
)

endpoint = model.deploy(
    machine_type="n1-standard-4",
    min_replica_count=1,
    max_replica_count=5,
    traffic_split={"0": 100}  # (1)!
)

# Predict
prediction = endpoint.predict(instances=[{"input": "data"}])
```

1. Traffic split enables A/B testing (multiple model versions)

```python
# Batch prediction (cost-effective)
batch_job = model.batch_predict(
    job_display_name="batch-prediction-job",
    gcs_source=["gs://my-bucket/input.jsonl"],
    gcs_destination_prefix="gs://my-bucket/output/",
    machine_type="n1-standard-4",
    starting_replica_count=1,
    max_replica_count=5
)

batch_job.wait()  # (1)!
```

1. Batch jobs cheaper than real-time (no endpoint costs)

```python
# Model Garden (pre-trained models)
from vertexai.preview.language_models import TextEmbeddingModel

# Use open-source models from Model Garden
embedding_model = TextEmbeddingModel.from_pretrained("textembedding-gecko@003")

embeddings = embedding_model.get_embeddings(["Text to embed"])
print(embeddings[0].values)  # (1)!

# Other Model Garden models: Llama 3, Mistral, Stable Diffusion, etc.
```

1. Embeddings for similarity search, clustering, RAG

```javascript
// Node.js SDK
const {VertexAI} = require('@google-cloud/vertexai');

const vertexAI = new VertexAI({
  project: 'my-project',
  location: 'us-central1'
});

const generativeModel = vertexAI.getGenerativeModel({
  model: 'gemini-2.0-flash-exp'
});

const result = await generativeModel.generateContent('Explain REST APIs');
console.log(result.response.text());  // (1)!
```

1. Node.js SDK mirrors Python API (consistent experience)

**Why this works:**

- Unified platform (training, deployment, monitoring)
- Native multimodal (text, image, video, audio)
- Grounding provides current information
- AutoML simplifies custom model training
- Model Garden offers 100+ pre-trained models