murtaza-nasir/speakr v0.8.8-alpha

v0.8.8 - Lightweight Docker Image

on GitHub

Lightweight Docker Image

New lite image tag that skips PyTorch and sentence-transformers for a dramatically smaller Docker image. All features work normally — only Inquire Mode's semantic search falls back to basic text search.

New Features

• Lite Image Tag - learnedmachine/speakr:lite at ~725MB vs ~4.4GB for the full image. Ideal for users who don't need semantic search or prefer a smaller footprint.

• Multi-Stage Dockerfile - Optimized build using static ffmpeg binaries (~150MB vs ~450MB from apt) and multi-stage compilation. Both latest and lite images benefit from smaller sizes.

• Improved Text Search Fallback - When embeddings are unavailable, Inquire Mode now uses stop word filtering, keyword-focused query enrichment, and match count ranking for much better results.

Image Tags

Usage

Pre-built (Docker Hub):

image: learnedmachine/speakr:lite

Building from source:

docker build --build-arg LIGHTWEIGHT=1 -t speakr:lite .

Bug Fixes

• Fixed SQLAlchemy lazy-load error in text search fallback (session detachment on TranscriptChunk.recording)

Compatibility

Fully backwards compatible with v0.8.x. The latest tag remains unchanged. The lite tag is a new addition.

Thanks to sakowicz for the suggestion (#214)