granite-sqlite-lfm 0.1.0

SQLite-backed local Ollama cascade orchestrator.

granite-sqlite-lfm

SQLite-backed local Ollama cascade orchestrator.

This project runs a local multi-stage model pipeline using Ollama and SQLite. It stores session state, extracted context variables, and conversation history in a local SQLite database, then routes user prompts through a lightweight extraction model, a worker model, and a fallback escalation model when needed.

Features

• SQLite-backed session and context storage
• Full-text searchable conversation history using SQLite FTS5
• Local Ollama model orchestration
• Stage 1 intent and variable extraction
• Stage 2 task execution
• Stage 3 fallback escalation
• Import-safe package structure
• CLI support with custom prompts and database paths

Requirements

• Python 3.9+

• Ollama installed and running locally

• Local Ollama models available:

  • lfm2.5:230m
  • ibm-granite/granite-4.0-h-350m-instruct
  • granite4.1:3b

Installation

Create and activate a virtual environment:

python -m venv .venv
source .venv/bin/activate

On Windows PowerShell:

.venv\Scripts\Activate.ps1

Install the package in editable development mode:

python -m pip install -e ".[dev]"

Usage

Run the package with the default demo prompts:

python -m granite_sqlite_lfm

Or use the console command:

granite-sqlite-lfm

Run a custom prompt:

granite-sqlite-lfm --prompt "Book a conference room for 15 attendees tomorrow at 3 PM."

Use a custom SQLite database path:

granite-sqlite-lfm --db-path ./local_context.db

Use a custom session ID:

granite-sqlite-lfm --session-id enterprise_user_session_101

Run multiple prompts in sequence:

granite-sqlite-lfm \
  --prompt "Book a conference room for 15 attendees tomorrow." \
  --prompt "Change the attendee count to 45."

Project Structure

granite-sqlite-lfm/
├── pyproject.toml
├── README.md
├── LICENSE
├── .gitignore
├── src/
│   └── granite_sqlite_lfm/
│       ├── __init__.py
│       ├── __main__.py
│       ├── cli.py
│       ├── config.py
│       ├── database.py
│       └── orchestrator.py
├── tests/
│   ├── test_database.py
│   └── test_orchestrator.py
└── examples/
    └── demo.py

Development

Install development dependencies:

python -m pip install -e ".[dev]"

Run tests:

pytest

Run Ruff linting:

ruff check .

Build the package:

python -m build

Check distribution artifacts:

twine check dist/*

Publishing

Before publishing, review the package metadata in pyproject.toml, including:

• package name
• author name
• license
• description
• supported Python version
• dependencies
• console script name

Then build and upload:

python -m build
twine check dist/*
twine upload dist/*

Notes

This project depends on a working local Ollama installation. The included tests should avoid real model calls and focus on package importability, SQLite behavior, JSON parsing, and CLI smoke behavior.

The default SQLite database file is:

cascade_context.db

This file is local runtime state and should not be committed to version control.