Datazone Agents

Datazone Agents are custom AI assistants that can interact with your data through natural language conversations. Each agent is configured with specific data sources, tools, and models to help users explore, analyze, and extract insights from their data.

What is an Agent?

An agent is an AI-powered assistant that:

• Understands natural language - Ask questions in any language
• Accesses your data - Queries datasets and views you’ve configured
• Uses tools - Executes SQL queries, runs Python code, creates charts, and more
• Provides insights - Analyzes data and delivers actionable answers

Unlike static dashboards, agents adapt to your questions in real-time and can handle complex, multi-step analysis automatically.

Creating an Agent

1. Navigate to your Project page
2. Click the Add button (+ icon)
3. Select Agent from the dropdown

The agent creation flow will guide you through 4 steps:

Step 1: Basic Information

Configure the basic settings:

• Name - A descriptive name for your agent (e.g., “Sales Analyzer”, “Customer Support Assistant”)
• Instructions - System prompt that defines the agent’s behavior, personality, and expertise
• Response Tone - How the agent communicates:
  • Casual - Informal, conversational style
  • Friendly - Warm and approachable
  • Technical - Precise and technical (default)
  • Educational - Explanatory and teaching-focused
  • Professional - Formal and business-like
• Response Length - How detailed responses should be:
  • Brief - Short, concise answers
  • Moderate - Balanced detail (default)
  • Detailed - Comprehensive, in-depth responses
• Language - The language your agent will use (supports multiple languages)

Step 2: Model Configuration

Select your AI model provider and model:

• Model Account - Choose from configured Model Accounts (OpenAI, Anthropic, AWS Bedrock)
• Model - Select the specific model (e.g., GPT-4, Claude 3.5 Sonnet, GPT-4o-mini)

Sampling settings such as temperature and max tokens are not set here. Add them under Advanced parameters on the agent’s Model tab after creating the agent. See Model Parameters.

Step 3: Data Sources

Choose which data sources the agent can access:

• Datasets - Direct access to raw datasets
• Views - Optimized, filtered data views
• Vectors - Semantic search on vectorized data for RAG (Retrieval Augmented Generation)

The agent will only be able to access data from these selected sources.

Step 4: Tools

Enable optional tools that extend what your agent can do:

• Python Code Executor - Execute Python code for calculations and analysis
• Chart Generator - Create visualizations (line, bar, pie charts)
• Web Search - Search the internet for external information
• Actions - Call custom functions you’ve deployed (e.g., send emails, trigger workflows)

Using Your Agent

Once created, you can start chatting with your agent:

1. Click on your agent from the agents list in the project page.
2. Click to Go to Agent to open the chat interface.
3. Then click to New Chat to start a fresh conversation
4. Ask questions in natural language (in any language you configured)
5. The agent will automatically use its tools to analyze data and provide answers

Agent Capabilities

Natural Language Queries

Ask questions in any language:

• “What were our top 5 customers last month?”
• “Show me the revenue trend for Q4”
• “Which products have declining sales?”
• “Compare this quarter to last year”

Semantic Search with Vectors (RAG)

When you add Vectors as data sources, agents can perform semantic similarity search to retrieve relevant context from your vectorized data. This enables:

• Context-aware responses - Agent finds relevant information based on meaning, not just keywords
• Document retrieval - Search through documents, knowledge bases, and unstructured data
• Enhanced accuracy - Provide agents with precise contextual information for better answers

The agent automatically decides when to use vector search based on the question and available data sources.

Action Tools

Agents can call custom actions you’ve deployed:

• Send notifications when certain conditions are met
• Trigger workflows based on data analysis
• Integrate with external services (Slack, email, APIs)
• Process or transform data using custom logic

The agent automatically decides when to use actions based on the conversation context. Learn more about creating actions.

Multi-Step Analysis

Agents can perform complex analysis automatically:

1. Execute SQL queries to fetch data
2. Run Python code for calculations
3. Generate visualizations automatically
4. Provide insights and recommendations

Context Awareness

Agents remember conversation history and can:

• Reference previous questions
• Build on earlier analysis
• Maintain context throughout the chat session

For facts that should carry over to later conversations, see Agent Memory.

Agent Memory

Assistants can remember across conversations. There are two kinds of memory, and each one has its own settings.

Your Memory

Memory holds what assistants should know about you, such as your role, the language you want answers in, or the metrics you care about. Assistants read it in every conversation and update it when you tell them something lasting about how you work.

• Limit - 2,000 characters
• Turn off - Switch off Let assistants remember you in Settings > Memory. While it is off, assistants neither read nor change it.
• Edit or clear - Edit the text in Settings > Memory and save. To clear it, delete the text and save; there is no separate delete.
• Visibility - Only you can see it. Organization admins cannot read it.

Memory describes you and your preferences, not data. Assistants do not store figures, query results, or personal identifiers in it.

Learned Instructions (Self Learner)

Learned instructions are rules that apply to everyone who uses an agent, such as “Use net revenue, not gross revenue.” They are separate from the agent’s Instructions. To turn them on, switch Self Learner on in the Learned Instructions card on the agent’s Overview tab. While it is on:

• The agent follows the learned instructions in every conversation.
• Users who can edit the agent can teach it new rules in chat. The agent records them when a rule should apply to every future conversation.
• Editors can also write the rules directly in the card, one short rule per line, up to 4,000 characters.

While Self Learner is off, the text is kept but not applied. Orion has the same switch and text in Project Settings > Orion, and intelligent apps have it in their LLM settings.

Best Practices

1. Clear Instructions - Write specific system instructions that define the agent’s expertise and behavior
2. Right Data Sources - Only grant access to relevant datasets/views
3. Appropriate Model - Choose models that balance cost and capability for your use case
4. Use Views - Configure views for faster queries and better data access control
5. Test Thoroughly - Ask various questions to ensure the agent understands your data correctly

Next Steps

• Chat with Your Agent
• Optimize Performance
• Configure Model Accounts