AI Data Analyst

Turn natural language questions into executable SQL against your local SQLite database. This project uses LangChain with local Ollama models and runs on the ultra-fast uv Python package manager.

Features

• Automatic schema extraction from amazon.db
• Natural language to SQL using local LLM (Ollama)
• Deterministic SQL generation (temperature=0)
• Easily switch models (e.g. DeepSeek-R1 vs QwenCoder)
• Simple Python API (get_data_from_database(prompt)) for integration
• Fast environment setup and dependency management via uv

Tech Stack

• Python 3.11+
• SQLite (amazon.db)
• SQLAlchemy for schema introspection
• LangChain Core + Community + Ollama integration
• Ollama local models (recommend: qwen2.5-coder:7b for speed)
• uv for dependency resolution, syncing, and running

Architecture

1. Extract schema using SQLAlchemy inspector
2. Feed schema + user question to an LLM prompt template
3. Model returns a SQL query (reasoning models may include <think> blocks – we strip or avoid them)
4. Execute SQL safely against amazon.db
5. Return results list

main.py
└─ extract_schema() -> text_to_sql() -> get_data_from_database()

Prerequisites

• Install Ollama: https://ollama.com/download
• Pull a model (choose one):
  • Fast coder model:

    ollama pull qwen2.5-coder:7b

  • Reasoning model (slower):

    ollama pull deepseek-r1:8b

• Ensure amazon.db exists in the project root.

Setup (UV)

No need to manually create a virtual environment—uv handles it.

# Install uv if not present
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install dependencies from pyproject.toml
uv sync

# Run a module or script
uv run python main.py

To add a new dependency:

uv add package_name

To upgrade dependencies:

uv lock --upgrade
uv sync

Usage

Python API example:

from main import get_data_from_database
results = get_data_from_database("Show the first 5 products")
print(results)

Direct script run (if you extend main.py with CLI later):

uv run python main.py

Switching Models

Edit main.py:

model = OllamaLLM(model="qwen2.5-coder:7b", temperature=0)

If you use DeepSeek-R1 models and see long delays, they generate hidden reasoning. Mitigations:

• Add: "Do not use tags." to system prompt
• Strip with regex: re.sub(r"<think>.*?</think>", "", raw, flags=re.DOTALL)
• Prefer a non-reasoning coder model for performance

Validating Generated SQL (Optional)

Before execution you can add a lightweight validator:

if not sql_query.lower().startswith("select"):
    raise ValueError("Only SELECT queries are allowed.")

Add more guards (block DROP/DELETE/UPDATE) for safety.

Performance Tips

• Use smaller local models (qwen2.5-coder:7b) for snappy responses
• Keep temperature=0 for deterministic output
• Cache schema: avoid recomputing extract_schema each call
• (Advanced) Stream model output and stop at first semicolon ; if the model over-explains

Troubleshooting

Issue	Cause	Fix
Long response time	Reasoning model generating chains	Switch to coder model
SQL errors (no such column)	Model hallucinated	Strengthen prompt, show schema clearly
Empty results	Query valid but data missing	Inspect amazon.db contents
Ollama connection error	Service not running	Run ollama serve or open app

Security Notes

Executing arbitrary LLM-generated SQL can be risky. Restrict to read-only queries and sanitize user inputs if you later interpolate values.

Contributing

1. Fork & branch
2. Make changes
3. Run uv sync && uv run python -m py_compile main.py
4. Submit PR

License

MIT

---

Built with local AI + fast Python tooling.