try-flockmtl

• FlockMTL/Ollama DuckDB Integration
  • Status
  • Why?
  • What?
  • How?
    • 1. Prerequisites
    • 2. Setup
      • a. Configure Models and Endpoint
      • b. Start Ollama
      • c. Register Models & Test
      • d. Manage with Justfile
    • 3. Troubleshooting
    • 4. Project Structure
    • 5. Extending
  • Example Usage
  • Compatibility Notes
  • License
  • Credits

FlockMTL/Ollama DuckDB Integration

Experiments with the FlockMTL DuckDB extension.

Status

There appears to be some issue with the FlockMTL extension and DuckDB 1.2.2, which is causing the GetAlterInfo not implemented for this type error. This might be due to a mismatch between the extension and the DuckDB version, however the documenation suggests that the extension is compatible with DuckDB 1.2.2.

Why?

Modern data workflows increasingly require local, private, and flexible access to large language models (LLMs) for analytics, summarisation, and automation. This project enables you to:

• Integrate local LLMs (via Ollama) directly with DuckDB using FlockMTL, so you can run LLM-powered SQL queries on your data.
• Automate and manage LLM model registration and configuration using Python and TOML, making your setup reproducible and easy to maintain.
• Debug and audit all SQL interactions for transparency and troubleshooting.

This is a common pattern that is emerging in the data ecosystem, where LLMs are used to augment SQL queries and data processing. For example:

• Note that the commercial MotherDuck offering offers built-in support for various LLMs via the prompt() method.
• Similarly, Google Sheets now has built-in support for the Gemini suite of LLMs via the =AI() function.

What?

This repository provides:

• (Configuration-driven, object-oriented Python script (flockmtl_manager.py) to:
  • Check Ollama endpoint health
  • Register Ollama models with FlockMTL in DuckDB
  • Run test completions via SQL
  • Log all SQL commands (with Loguru) and export them to flockmtl.sql
• Sample TOML config (flockmtl.toml) for easy model and endpoint management
• Justfile for convenient CLI automation of common tasks (running scripts, managing Ollama models, etc.)

How?

1. Prerequisites

• Python 3.13+
• DuckDB installed (CLI or Python package)
• Ollama installed and running (see Ollama docs)
• FlockMTL DuckDB extension installed
• Python dependencies:

uv add duckdb httpx loguru

• (Optional) Just for task automation

2. Setup

a. Configure Models and Endpoint

Edit flockmtl.toml to specify your Ollama endpoint and models. For example:

[ollama]
api_url = "http://127.0.0.1:11434"

[[models]]
name = "mixtral_local"
ollama_name = "mixtral:latest"
context_window = 128000
max_output_tokens = 2048

[[models]]
name = "llama2_local"
ollama_name = "llama2:latest"
context_window = 4096
max_output_tokens = 1024

b. Start Ollama

ollama serve

or use the Justfile recipe:

just ollama-serve

c. Register Models & Test

Run the Python manager script:

python flockmtl_manager.py

This will:

• Check if the Ollama endpoint is available
• Register the Ollama API secret with FlockMTL
• Register each model from the TOML file
• Run a test LLM completion via SQL
• Log all SQL statements to flockmtl.log and flockmtl.sql

d. Manage with Justfile

Use the included Justfile for tasks like:

just ollama-list         # List local Ollama models
just ollama-pull llama2  # Pull a model
just run                 # Run the manager script

3. Troubleshooting

• Extension errors: Ensure the FlockMTL extension matches your DuckDB version and platform. Use INSTALL flockmtl; LOAD flockmtl; in DuckDB, or see DuckDB extension docs.
• Hanging on test completion:
  • Make sure Ollama is running and the specified model is available.
  • Test the model directly: ollama run llama2 --prompt "Hello"
• SQL errors or debugging:
  • Check flockmtl.sql for all executed SQL statements.
  • Review flockmtl.log for detailed logs.

4. Project Structure

.
├── flockmtl_manager.py   # Main Python script (OO, config-driven)
├── flockmtl.toml         # Model & endpoint configuration
├── flockmtl.sql          # All executed SQL commands (for debugging)
├── flockmtl.log          # Loguru logs (SQL and actions)
├── Justfile              # Task automation commands

5. Extending

• Add more models to flockmtl.toml as needed.
• Customise or extend the Python script for advanced workflows (e.g., model removal, batch prompts).
• Use the SQL log for bug reports or reproducibility.

---

Example Usage

# List available tasks
just

# Start Ollama server
just ollama-serve

# Pull a new model
just ollama-pull llama2

# Run the integration script
just run

---

Compatibility Notes

• DuckDB and FlockMTL extension versions must be compatible. If you see errors about GetAlterInfo not implemented for this type, update your FlockMTL extension to match your DuckDB version.
• Ollama models must be pulled before use.
• Python 3.13+ is required for built-in TOML parsing.

---

License

Apache-2.0 License

This project is licensed under the Apache-2.0 License. See the LICENSE file for details.

---

Credits

• FlockMTL
• Ollama
• DuckDB
• Loguru
• httpx

---

Questions or issues?

Open an issue or contact [email protected].