OSS Data Analyst

An AI data analyst agent that explores a semantic layer in a sandbox environment to answer natural language questions with SQL.

Overview

OSS Data Analyst uses a sandboxed exploration approach: instead of hardcoding schema knowledge into prompts, the agent is given shell access to a sandbox containing your semantic layer files. It discovers the schema dynamically using cat, grep, and ls commands, then builds and executes SQL queries based on what it finds.

This architecture means the agent can:

• Adapt to any schema without prompt changes
• Explore relationships between entities naturally
• Handle schema updates without redeployment
• Reason about data the same way a human analyst would

How It Works

1. Sandbox Creation - A Vercel Sandbox is spun up and populated with your semantic layer YAML files
2. Schema Exploration - The agent uses shell commands to browse the catalog and entity definitions
3. Query Building - Based on discovered schema, the agent constructs SQL queries
4. Execution - Queries run against your SQLite database
5. Reporting - Results are formatted with a narrative explanation

1
User Question
2
     ↓
3
┌─────────────────────────────────────┐
4
│           Vercel Sandbox            │
5
│  ┌─────────────────────────────┐   │
6
│  │  semantic/                   │   │
7
│  │  ├── catalog.yml            │   │
8
│  │  └── entities/              │   │
9
│  │      ├── companies.yml      │   │
10
│  │      ├── people.yml         │   │
11
│  │      └── accounts.yml       │   │
12
│  └─────────────────────────────┘   │
13
│                                     │
14
│  Agent explores with:               │
15
│  • cat semantic/catalog.yml         │
16
│  • grep -r "keyword" semantic/      │
17
│  • cat semantic/entities/*.yml      │
18
└─────────────────────────────────────┘
19
     ↓
20
SQL Query → Database → Results → Narrative

Quick Start

Prerequisites

• Node.js 20+
• pnpm
• Vercel AI Gateway API key

Installation

1
git clone https://github.com/vercel-labs/oss-data-analyst.git
2
cd oss-data-analyst
3
pnpm install

Configuration

1
cp env.local.example .env.local

Add your Vercel AI Gateway key to .env.local.

Initialize Database

1
pnpm initDatabase

Creates a SQLite database with sample data (Companies, People, Accounts).

Run

1
pnpm dev

Open http://localhost:3000

Semantic Layer

The semantic layer lives in src/semantic/ and defines your data model:

1
src/semantic/
2
├── catalog.yml           # Entity index with descriptions
3
└── entities/
4
    ├── companies.yml     # Company entity definition
5
    ├── people.yml        # People entity definition
6
    └── accounts.yml      # Accounts entity definition

Each entity YAML includes:

• sql_table_name - The underlying table
• fields - Available columns with SQL expressions
• joins - Relationships to other entities
• Example questions the entity can answer

The agent reads these files at runtime to understand your schema.

Example Questions

• "How many companies are in the Technology industry?"
• "What is the average salary by department?"
• "Show me the top 5 accounts by monthly value"
• "Which companies have the most employees?"

Architecture

Stack: Next.js, Vercel AI SDK, Vercel Sandbox, SQLite

Key Files:

• src/lib/agent.ts - Agent definition and system prompt
• src/lib/tools/sandbox.ts - Sandbox creation with semantic files
• src/lib/tools/shell.ts - Shell command tool for exploration
• src/lib/tools/execute-sqlite.ts - SQL execution tool

Adding Your Own Schema

1. Add entity YAML files to src/semantic/entities/
2. Update src/semantic/catalog.yml with the new entity
3. The agent will automatically discover and use the new schema

No code changes required—the sandbox approach means schema changes are picked up at runtime.

Troubleshooting

Database Not Found

1
pnpm initDatabase

Build Errors

1
pnpm type-check