v8.75.0 - Lightweight ONNX Quality Scoring

🎯 90% Installation Size Reduction

Major achievement: Installation shrinks from 7.7GB to 805MB while maintaining full quality scoring functionality.

Key Features

Lightweight ONNX Setup (PR #337)

• Problem Solved: transformers package added 6.9GB of dependencies (torch, tensorflow, etc.)
• Solution: Use tokenizers package directly instead of full transformers stack
• Benefits:
  • 90% disk space reduction: 7.7GB → 805MB
  • Faster installation: <2 min (vs 10-15 min)
  • Same quality scoring performance (nvidia-quality-classifier-deberta ONNX model)
  • Conditional dependency loading - only install what you use
  • No runtime performance impact

Implementation Details

• Modified onnx_ranker.py to use tokenizers directly
• Added tokenizers as optional dependency in pyproject.toml
• Updated embedding service with graceful fallback (tokenizers/transformers)
• Fixed quality_provider metadata access bug in async_scorer.py
• 15 comprehensive integration tests (487 lines)

Documentation & Tooling

• Complete setup guide: docs/LIGHTWEIGHT_ONNX_SETUP.md
• Automated setup script: scripts/setup-lightweight.sh
• One-command installation for lightweight mode

Bug Fixes

Multi-Protocol Port Detection (Issue #341)

• Problem: Update script failed on Linux systems without lsof
• Root Cause: Script used lsof exclusively, health checks only tried HTTP
• Solution:
  • Port detection fallback chain: lsof → ss → netstat → ps
  • Health checks support both HTTP and HTTPS with automatic fallback
  • Tested on Arch Linux with ss-only environment

Installation

New Users (lightweight installation):

pip install mcp-memory-service
# Lightweight ONNX setup (805MB total)
bash scripts/setup-lightweight.sh

Existing Users (upgrade):

pip install --upgrade mcp-memory-service

Full Installation (with transformers for embeddings):

pip install mcp-memory-service[full]

What's Next

• Continue optimizing installation size
• Explore additional lightweight ML backends
• Performance benchmarking with tokenizers-only setup

See CHANGELOG.md for full version history.

🚀 Upgrade now for a dramatically leaner installation!