demo-t2sql

A text-to-SQL workflow package. Accepts a natural language question, generates SQL via an LLM (Ollama/llama3.2), runs it against a SQLite database (Chinook), and streams back the SQL, result rows, and a plain-English summary.

Designed to be called by an external platform.

---

How it works

question → ChromaDB (RAG context) → Ollama (SQL generation) → SQLite (execution) → summary

1. ChromaDB retrieves relevant schema/examples as context
2. llama3.2 generates SQL from that context
3. SQL runs against data/Chinook.sqlite
4. llama3.2 summarises the results in plain English

---

Setup

uv sync
ollama pull llama3.2

The ChromaDB knowledge base is pre-populated. If you need to rebuild it, run data/setup_knowledge.py once.

---

Entry points

File	Style	Use when
src/default.py	Linear, simple	Standard use
src/workflow.py	pydantic_graph nodes	Need branching/conditional logic

Both expose the same signature:

async def run_workflow(request: QueryRequestModel, llm_id: str = "default", **kwargs)
    -> AsyncGenerator[ResponseEventToken | ResponseEventNode | ResponseEventProgress, None]

---

Key files to edit

What you want to change	File
Switch LLM (Ollama ↔ OpenAI)	src/tools/vanna_client.py
Customise the summary prompt	src/tools/vanna_client.py — uncomment generate_summary override
Add/change pipeline steps	src/default.py
Add graph nodes / branching	src/nodes/ + src/workflow.py
Retrain the knowledge base	data/setup_knowledge.py

---

Switching to OpenAI

In src/tools/vanna_client.py, uncomment the OpenAI import and class, comment out the Ollama ones, and set your key:

# "model": "gpt-4o",
# "api_key": os.environ["OPENAI_API_KEY"],