Curie — Local Text2SQL for PostgreSQL and SQLite

Curie is a VS Code extension that turns natural-language questions into PostgreSQL or SQLite queries. It runs a fine-tuned Qwen model locally, so schema context and query generation stay on your machine.

Curie was developed as a university project for local, understandable Text2SQL workflows.

Academic use and licensing

Curie was developed as a non-commercial student project for a Technical University of Munich (TUM) module and is intended for research and evaluation.

Improved using Qwen. The selectable Curie compact (0.5B) and Curie medium (1.5B) adapters are based on models licensed under the Apache License 2.0. The default Curie standard (3B) adapter is subject to the Qwen Research License Agreement. Commercial use of the 3B model or adapter requires separate permission from Alibaba Cloud. The adapters were trained using the Spider dataset, licensed under CC BY-SA 4.0.

What you can do

• Connect to a PostgreSQL database or select a local SQLite file from the Curie sidebar.
• Inspect tables, columns, primary keys, and foreign-key relationships.
• Switch between the bundled Curie compact (0.5B), Curie medium (1.5B), and Curie standard (3B) adapters.
• Select the tables that should be included in the model context.
• View the selected schema as a table list or Mermaid ER diagram.
• Generate SQL for the selected database from questions written in natural language.
• Review and edit generated SQL before running it.
• Run a query and view up to 100 returned rows directly in VS Code.
• Copy query results as tab-separated values for use in a spreadsheet.

For example, ask:

Which customers placed the most orders last month?

How it works

1. Curie starts a managed, isolated Python environment on your computer.
2. It installs the curie-backend package when needed and starts the local API on 127.0.0.1.
3. You load and select the relevant PostgreSQL or SQLite schema.
4. Curie combines your question with that schema and generates SQL using the selected local Qwen model.
5. You review the query and may run it against the configured database.

The initial setup can take several minutes: the Python backend, its dependencies, and the public Qwen base model may need to be downloaded. Subsequent starts use the locally stored environment and model cache.

The backend package contains roughly 220 MB of uncompressed adapter data. Each selected public Qwen base model is downloaded separately on first use, so ensure that additional disk space is available for the Hugging Face model cache.

Requirements

• VS Code 1.105 or newer
• Python 3.10 or newer
• A reachable PostgreSQL database or local SQLite database file
• Sufficient local memory and disk space for the selected Qwen model
• Internet access for the initial backend installation and model download

Get started

1. Install the extension and open Curie from the Activity Bar.
2. Wait until the backend status shows Running. If prompted, choose Install backend.
3. In Database, choose PostgreSQL and enter its connection details, or choose SQLite and select a local database file. Then select Load schema.
4. Select the tables relevant to your question.
5. Enter a question and select Generate SQL.
6. Review or edit the generated statement.
7. Keep Read-only transaction enabled and select Run SQL when ready.

Privacy and safety

• Model inference runs locally; Curie does not use a hosted inference service.
• The backend listens only on 127.0.0.1.
• Database credentials and the selected SQLite path are stored with VS Code SecretStorage, not in the saved webview state.
• Curie does not collect telemetry.
• SQL execution accepts one statement per request and returns at most 100 rows.
• Read-only execution is enabled by default for PostgreSQL and SQLite.

Generated SQL can be incorrect or unsafe for your data. Always review it before execution. Disable read-only mode only when you intentionally want to run a statement that changes the database.

Commands

Open the Command Palette and search for Curie:

• Curie: Open Text2SQL Assistant
• Curie: Start Backend Server
• Curie: Stop Backend Server

Settings

Troubleshooting

Python was not found

Set dbAssistant.backendPythonPath to a Python 3.10+ interpreter, for example:

/usr/local/bin/python3

The backend cannot start

Open View → Output, select Curie Backend, and inspect the installation or model-loading log. If the configured port is already in use, select a different dbAssistant.backendPort and restart the backend.

Recover an interrupted backend installation

First select Start backend again and repeat the installation. If the managed Python environment itself remains unusable, close all VS Code windows before recreating it with the commands below.

These paths apply to the standard VS Code installation. VS Code Insiders and other distributions use a different user-data directory.

macOS

CURIE_ENV="$HOME/Library/Application Support/Code/User/globalStorage/curie.curie-text2sql/backend-venv"
python3 -m venv --clear "$CURIE_ENV"
"$CURIE_ENV/bin/python" -m pip install --upgrade --index-url https://gitlab.lrz.de/api/v4/projects/256925/packages/pypi/simple curie-backend==0.2.8

Linux

CURIE_ENV="$HOME/.config/Code/User/globalStorage/curie.curie-text2sql/backend-venv"
python3 -m venv --clear "$CURIE_ENV"
"$CURIE_ENV/bin/python" -m pip install --upgrade --index-url https://gitlab.lrz.de/api/v4/projects/256925/packages/pypi/simple curie-backend==0.2.8

Windows PowerShell

$CurieEnv = Join-Path $HOME ".curie\backend-venv"
py -m venv --clear $CurieEnv
& (Join-Path $CurieEnv "Scripts\python.exe") -m pip install --upgrade --index-url https://gitlab.lrz.de/api/v4/projects/256925/packages/pypi/simple curie-backend==0.2.8

Reopen VS Code and select Start backend. This recreates only Curie's managed Python environment; saved database credentials and the separate Hugging Face model cache are not removed.

The database connection fails

Check the host, port, database name, username, password, and SSL mode. Also confirm that PostgreSQL accepts connections from the computer running VS Code. For SQLite, confirm that the selected local file still exists and is readable.

The first start takes a long time

This is expected while Curie installs the backend dependencies or downloads the base model. Keep VS Code open and check Curie Backend output for progress.