DeepDive – Universal Data Agent

DeepDive transforms natural language into database queries, generates visualizations, and learns from user corrections to improve over time.

Quick Start

# Setup (creates .deepdive/ directory)
@deepdive init

# Configure database in .deepdive/.env
DATABASE_URL=postgresql://user:pass@localhost:5432/db

# Query data
@deepdive query "show top 10 customers by revenue"

# Visualize schema
@deepdive visualize schema

# Create chart
@deepdive chart "monthly revenue over last 6 months"

Core Commands

Database Connection

• @deepdive init – Initialize .deepdive/ directory with .env template
• @deepdive connect <type> – Setup connection (postgres|mysql|sqlite|bigquery|snowflake)

Natural Language Queries

• @deepdive query "<question>" – Convert question to SQL and execute
• @deepdive preview "<query>" – Show results as markdown table (limit 100 rows)

Visualization

• @deepdive visualize schema – Generate ERD diagram (.deepdive/diagrams/)
• @deepdive visualize lineage – Show table relationships
• @deepdive chart "<question>" – Generate Vega-Lite chart (.deepdive/charts/)

Learning & Safety

• @deepdive learn – View/update learned corrections
• @deepdive history – Show recent queries
• @deepdive safe-mode [on|off] – Require confirmation for writes (default: on)

Project Structure

DeepDive creates and manages:

.deepdive/
├── .env                      # Database credentials (user-managed)
├── memory.json               # Learned corrections (per-project)
├── diagrams/                 # Generated .mmd files
│   ├── schema-YYYYMMDD.mmd
│   └── erd-YYYYMMDD.mmd
├── charts/                   # Generated .png/.svg files
│   └── chart-XXX.png
└── queries.log               # Query history

Supported Databases

• PostgreSQL – Full support with advanced features
• MySQL – Standard SQL support
• SQLite – File-based, perfect for local/dev
• BigQuery – Google Cloud, large-scale analytics
• Snowflake – Cloud data warehouse
• Redshift – AWS analytics

Reference Documentation

Read these files based on the task:

• Database Connections: references/connectors.md
• Natural Language Queries: references/nl-to-sql.md
• Schema Introspection: references/schema-introspection.md
• Mermaid Visualization: references/mermaid-viz.md
• Vega-Lite Charts: references/vega-charts.md
• User Learning: references/user-learning.md
• Write Protection: references/write-protection.md
• Examples: references/examples.md

Usage Patterns

Data Exploration

User: "What tables are in this database?"
→ @deepdive schema introspection

User: "Show me the customer table structure"
→ @deepdive schema customers

Querying

User: "Which customers bought something last month?"
→ Natural language → SQL → Execute → Results

User: "Chart monthly revenue"
→ Query → Vega-Lite spec → .deepdive/charts/revenue.png

Visualization

User: "Visualize the database schema"
→ Mermaid ERD → .deepdive/diagrams/schema.mmd → Open browser

User: "Show relationships between tables"
→ Foreign key analysis → Lineage diagram

Key Principles

1. Environment Variables: All credentials in .deepdive/.env (never hardcoded)
2. Write Protection: INSERT/UPDATE/DELETE require explicit confirmation (unless safe-mode off)
3. Learning: Corrections stored in .deepdive/memory.json and applied to future queries
4. Project Scope: Each project has isolated memory, diagrams, and charts
5. Static Outputs: All visualizations are files (.mmd, .png) for version control

Examples

RevOps Scenario

@deepdive connect postgres
@deepdive query "qualified opportunities by stage this quarter"
@deepdive chart "conversion funnel"
@deepdive visualize lineage

Video Production Scenario

@deepdive connect sqlite  # For document analysis
@deepdive query "projects due this week from documents table"
@deepdive chart "project timeline"

Scripts

Python scripts for reliable operations:

• scripts/generate_mermaid.py – Generate schema/ERD diagrams
• scripts/generate_chart.py – Create Vega-Lite charts
• scripts/validate_query.py – SQL safety validation

Execute scripts rather than rewriting code for deterministic results.

Safety & Privacy

• Read-only by default for exploration
• Write operations require confirmation
• Credentials never logged or shared
• All data stays local (no cloud API calls)
• Query history stored locally only