DataSmith AI - Intelligent Dataset Generation Platform

Status: Phase 0 Complete ✅ - Foundations Ready

---

Overview

DataSmith AI is a systematic dataset generation platform built with strong foundations:

• Data Contract: Universal JSON schema for the entire pipeline
• LLM Wrapper: Reusable Ollama interface with proper error handling
• Orchestrator: Pipeline runner with save/resume capability

Phase 0 provides the infrastructure that all future agents will use.

---

Project Structure

DataSmith_AI/
├── core/                    # Phase 0 foundations
│   ├── __init__.py
│   ├── dataset_spec.py      # Data contract
│   ├── llm_wrapper.py       # LLM wrapper
│   └── orchestrator.py      # Pipeline orchestrator
├── tests/                   # Test suite
│   ├── __init__.py
│   ├── test_dataset_spec.py
│   ├── test_llm_wrapper.py
│   └── test_orchestrator.py
├── venv/                    # Virtual environment
├── .gitignore
├── requirements.txt
├── run.bat
└── README.md                # This file

---

Installation

# 1. Create virtual environment
python -m venv venv

# 2. Activate
.\venv\Scripts\activate

# 3. Install dependencies
pip install -r requirements.txt

# 4. Install Ollama and pull model
# Download from: https://ollama.com/download
ollama pull qwen2.5:3b

---

Running Tests

# Activate virtual environment
.\venv\Scripts\activate

# Run all tests
python -m pytest tests/ -v

# Run specific test file
python -m pytest tests/test_dataset_spec.py -v

Expected: 19/19 tests passing (4 skipped without Ollama)

---

Phase 0 Components

1. Data Contract (core/dataset_spec.py)

• Universal JSON schema for entire pipeline
• Provenance tracking (source_url on every record)
• Confidence scores throughout
• Execution history for save/resume
• Tests: 8/8 passing

2. LLM Wrapper (core/llm_wrapper.py)

• Reusable Ollama interface
• JSON-only output with validation
• Total timeout limit (60s default)
• Maximum 1 retry with error feedback
• Connection error detection
• Tests: 5/5 passing (4 skipped)

3. Orchestrator (core/orchestrator.py)

• Runs agents in sequence
• Saves spec after each agent
• Resume from any step
• Skip conditions and error handling
• Tests: 6/6 passing

---

Next Steps: Phase 1

Phase 1 will build agents using these foundations:

1. Requirement Understanding Agent

   • Uses LLMWrapper
   • Writes to spec.requirement
   • Handles user input

2. Domain Detection Agent

   • Reads spec.requirement
   • Adds domain classification
   • Uses shared data contract

All future agents will follow the same pattern.

---

Key Principles

1. ✅ No Code Duplication - All agents use LLMWrapper
2. ✅ Unified Data Contract - One spec flows through pipeline
3. ✅ Provenance - source_url on all data
4. ✅ Testable - Comprehensive test coverage
5. ✅ Resumable - Save/resume at any step
6. ✅ Generic - Works for any domain

---

Dependencies

• Python 3.8+
• Ollama (local LLM)
• qwen2.5:3b model
• pydantic (data validation)
• pytest (testing)

---

Documentation

• core/dataset_spec.py - Data contract with inline docs
• core/llm_wrapper.py - LLM wrapper with inline docs
• core/orchestrator.py - Orchestrator with inline docs
• LLM_WRAPPER_IMPROVEMENTS.md - Recent LLM wrapper improvements

---

Last Updated: 2026-10-03
Phase: 0 Complete, 1 Pending