> ## Documentation Index
> Fetch the complete documentation index at: https://docs.datazone.co/llms.txt
> Use this file to discover all available pages before exploring further.

# Best Practices

> Optimize your agents for better performance, accuracy, and cost efficiency

# Best Practices

Follow these guidelines to build **effective, efficient agents** that deliver **accurate insights** while **managing costs**.

## Model Selection

### Choosing the Right Model

Different models excel at different tasks:

| Model Type            | Best For                                 | Cost   | Speed  |
| --------------------- | ---------------------------------------- | ------ | ------ |
| **GPT-4**             | **Complex reasoning**, detailed analysis | High   | Slower |
| **GPT-4o**            | **Balanced performance**, general use    | Medium | Medium |
| **GPT-4o-mini**       | **Simple queries**, high volume          | Low    | Fast   |
| **Claude 3.5 Sonnet** | **Code generation**, structured data     | Medium | Medium |
| **Claude 3 Haiku**    | **Quick answers**, basic queries         | Low    | Fast   |

### Use Case Guidelines

**Complex Analysis** → GPT-4 or Claude 3.5 Sonnet

* **Multi-step reasoning** required
* **Deep data analysis**
* **Code generation** needs

**General Use** → GPT-4o or Claude 3.5 Sonnet

* **Balanced performance**
* Most common questions
* **Good for production** agents

**High Volume** → GPT-4o-mini or Claude 3 Haiku

* **Simple queries**
* **Cost-sensitive** applications
* **Fast response times** needed

<Tip>
  Start with a **mid-tier model** (GPT-4o or Claude 3.5 Sonnet) and adjust based on actual usage patterns.
</Tip>

## Model Parameters

### Temperature

**Temperature** controls how **creative** or **focused** the agent's responses are:

<Frame>
  <img src="https://mintcdn.com/datazone/-RitKVuJcAfZzcZR/images/light/agent/agent-llm-temperature.jpeg?fit=max&auto=format&n=-RitKVuJcAfZzcZR&q=85&s=a89e531eb17d1c51ada9a3e18a7da879" alt="Temperature Scale" width="1888" height="560" data-path="images/light/agent/agent-llm-temperature.jpeg" />
</Frame>

* **0.0 - 0.3** (Focused)
  * **Consistent, deterministic** answers
  * Best for: **Data queries**, factual analysis
  * Use when **accuracy is critical**

* **0.4 - 0.7** (Balanced)
  * Mix of **consistency and creativity**
  * Best for: **General purpose** agents
  * **Recommended default**

* **0.8 - 1.0** (Creative)
  * **Varied, exploratory** responses
  * Best for: Brainstorming, recommendations
  * Use **sparingly for data agents**

<Note>
  For most data analysis agents, keep temperature between **0.2 and 0.5** for reliable results.
</Note>

### Max Tokens

**Max Tokens** controls the **maximum response length**:

* **1,000 - 2,000** - **Short, focused** answers
* **2,000 - 4,000** - **Standard responses** (recommended)
* **4,000+** - **Detailed analysis** and long explanations

**Higher limits = more cost**. Set based on expected answer complexity.

## Data Source Optimization

### Use Views Instead of Datasets

**Why Views Are Better:**

* ✅ **Faster query execution**
* ✅ **Pre-filtered**, relevant data only
* ✅ **Better security** (control data access)
* ✅ **Lower token usage**
* ✅ **More accurate** agent responses

**Example:**

```
Instead of: Full "orders" dataset (millions of rows)
Create: "recent_orders" view (last 90 days, key columns)
```

The agent queries **faster** and gets **relevant data immediately**.

<Warning>
  **Always use Views instead of raw datasets** for better performance and cost efficiency.
</Warning>

### Limit Data Sources

Only grant access to **necessary datasets/views**:

**Too Broad:**

* Sales data
* HR data
* Marketing data
* Finance data

**Focused:**

* Sales summary view
* Revenue trends view

**Fewer sources = faster decisions + lower costs**.

### Optimize View Definitions

Create views with:

* **Only necessary columns**
* **Pre-aggregated data** where possible
* **Relevant date ranges**
* **Indexed fields**

## Agent Instructions

### Write Clear System Prompts

**Good Instructions:**

```
You are a sales analytics expert. Help users analyze:
- Revenue trends and forecasting
- Customer behavior and segmentation
- Product performance metrics

Always provide specific numbers and dates. When showing
trends, create visualizations. Be concise and actionable.
```

**Poor Instructions:**

```
You are helpful. Answer questions about data.
```

### Define Scope

Tell the agent:

* **What data** it has access to
* **What questions** it should handle
* **What format** responses should take
* Any **business rules** to follow

### Include Examples

Provide **specific examples** in instructions:

```
When asked about "top customers," show:
1. Customer name
2. Total revenue
3. Number of orders
4. Comparison to average

Example format:
"Top 5 customers by revenue:
1. Acme Corp - $150K (45 orders) - 3x average
..."
```

## Cost Optimization

### Control Token Usage

**Input Tokens:**

* Keep **system instructions concise**
* Limit **conversation history** length
* Use **focused data sources**

**Output Tokens:**

* Set **appropriate max\_tokens**
* Request **concise answers** when possible
* Avoid asking for **repeated information**

### Monitor Spending

* Review **token usage per conversation**
* **Track costs by agent**
* Set up **alerts for high usage**
* Regularly **audit agent performance**

### Batch Similar Questions

If running automated analysis:

* **Group related queries**
* **Reuse context** where possible
* **Cache frequently** accessed data

## Agent Configuration

### Tool Selection

Enable **only what you need**:

**Optional Tools:**

* **Chart Generator** - For data visualizations (recommended for most agents)
* **Python Executor** - For complex calculations and advanced analytics
* **Web Search** - For external context and current information

<Note>
  **Query Executor** is enabled by default and cannot be disabled.
</Note>

### Data Access Control

* Grant **minimum necessary access**
* Use **views to restrict data**
* **Separate agents** by use case
* **Review permissions** regularly

### Regular Maintenance

* **Update instructions** based on user feedback
* **Refine data sources** as needs change
* **Adjust model selection** for cost/performance
* **Archive unused agents**

## Query Performance

### Faster Queries

1. **Use Views** - **Pre-filtered, optimized data**
2. **Limit Result Sets** - Ask for **"top 10"** not "all"
3. **Specific Time Ranges** - **"last month"** not "all time"
4. **Indexed Columns** - Ensure views use **indexed fields**

### Example Optimizations

**Slow:**
"Show all customer transactions"

**Fast:**
"Show top 20 customers by revenue this quarter"

## Testing & Validation

### Test Common Questions

Before deploying:

* Ask **typical user questions**
* **Verify data accuracy**
* Check **tool usage patterns**
* Measure **response times**

### Validate Answers

* Compare agent responses to **known results**
* **Verify calculations** manually
* Check **chart accuracy**
* Test **edge cases**

### Iterate

* Gather **user feedback**
* **Refine instructions**
* **Adjust parameters**
* **Optimize data sources**

## Security

### Data Access

* Only grant **necessary permissions**
* Use **views to limit sensitive data**
* **Review agent access** regularly
* **Audit query logs**

### Credentials

* **Secure model account** API keys
* **Rotate credentials** periodically
* **Monitor for unusual usage**
* Implement **access controls**

## Next Steps

* [Create Your First Agent](/reference/agents/overview)
* [Configure Model Accounts](/reference/development/model-accounts)
* [Set Up Data Views](/reference/integration/views)
* [Chat with Your Agent](/reference/agents/chat)