# Gemini 3.0 Pro Preview

We have observed that **Gemini 3.0 Pro Preview** provisioned by VertexAI can exhibit **significantly long response times**. This behavior is caused by **load conditions on Google’s Vertex AI infrastructure**, which can not be influenced by Unique.

# **How to Use Gemini 3 Pro Preview with UniqueAI via LiteLLM**  
Gemini 3 Pro Preview supports a **large 1M-token context window**, but it also comes with important constraints:

- **Quota limit is close to the full context size**  
- **Full-window LLM calls are slow**  
- **Large-token executions become expensive very quickly**

To ensure fast, stable, and cost-efficient usage with UniqueAI, we recommend the following configuration.

* * *

## **1. Limit Returned Chunks per Tool Call (max 100 chunks)**  
To avoid hitting token limits, long runtimes, and high costs, internal knowledge search tool calls should be restricted to return **max 100 chunks**.

### **Where to configure it**
In the **Internal Knowledge Search configuration**, set:

```
limit = 100
```

* * *

## **2. Set Gemini “Thinking Level” to Low (for better speed)**  
Gemini 3 Pro Preview supports two reasoning [modes](https://ai.google.dev/gemini-api/docs/openai):

- **high** (slower, higher cost)  
- **low** (faster, lower cost)

For most flows in UniqueAI, **low** provides more than enough reasoning quality while dramatically improving speed and reducing token usage.

### **Where to configure it**
In the **Loop Agent configuration** under **Experimental → Additional LLM Options**, add the `reasoning_effort` as low.

Config path: `Advanced Settings` → `Loop Agent` → `Experimental` → `Additional Llm Options`

* * *

## **Recommended Configuration Summary**  
| **Setting** | **Location** | **Recommended Value** |
| --- | --- | --- |
| Max return chunks per tool call | Internal Knowledge Search → limit | **100** |
| Reasoning level | Loop Agent → Experimental → Additional Llm Options | **reasoning_effort: “low”** |