Deep Research

The Deep Research agent is an expert analytics assistant designed to tackle complex, multi-faceted questions. It breaks down complex requests into systematic research plans, executes them step-by-step, and delivers comprehensive, data-driven summaries.

How it works

The agent follows this workflow:

1. Analyzes and decomposes the user’s research request
2. Discovers relevant data products if none are specified
3. Retrieves schemas to understand available data and relationships
4. Creates a structured research plan with business-friendly language (or uses a provided plan)
5. Executes each research step independently, running SQL queries to extract data
6. Documents findings in plain language as it progresses
7. Synthesizes all findings into a comprehensive final summary with actionable insights and specific metrics

Input parameters

Required:

• message (string): The complex question or research request

Output format

The agent produces a series of thinking, tool call, tool return, and text blocks as it works through the user request. The final message, assuming no errors, is a json object containing:

• summary (string): A comprehensive final summary addressing the original question
• plan (object): The research plan used (saveable for future use)
  • plan_title (string): The title of the research plan
  • plan_steps (array): List of research steps, each containing:
    • section_header (string): The main goal for this research step
    • instructions (array of strings): The list of analytics questions for the step
  • plan_goal (string, optional): The main research goal

Available tools

The agent has access to three tools:

List available data products

Discovers relevant data products based on the user’s question. Returns up to 5 data products by default using semantic search.

Get data schema

Retrieves schema information for data products. Supports optional semantic search of sample values.

SQL execution

Executes SQL queries against data products to extract insights and data.

Behavior notes

• No clarification questions: The agent uses best judgment to interpret vague requests and provides complete answers
• Data-driven focus: Prioritizes calculations, aggregations, trends, and specific metrics over metadata
• Business-friendly language: Avoids SQL, tool names, and technical identifiers in user-facing text
• Concurrent execution: Invokes multiple independent tools simultaneously for efficiency
• Critical thinking: Carefully analyzes tool results to ensure high-quality, accurate answers
• Comprehensive results: Always aims to fully address the request, even if information is incomplete