eloquent-notes

Eloquent Notes is a Linux system-tray utility for quick offline dictation into Obsidian. It runs in the background, records short voice clips, and processes them entirely on the local machine using a Gemma 4 model served through Ollama. Workflow: click the tray icon (or trigger a global shortcut / CLI toggle) to start recording, click again to stop. Audio is captured in memory, converted to WAV bytes, and sent through a three-phase LLM pipeline: transcription, rewriting into first-person note prose with a title, and classification producing a note type, wikilinks and tags. The result is written to the configured Obsidian vault, either as a standalone file or appended to a daily note. Notable capabilities described in the README: - System tray UX with dynamically drawn state icons (idle, recording, processing) and a context menu for configuration, reload and quit. - CLI commands: toggle recording, open the configuration GUI, and install an autostart entry. The toggle command communicates with the running daemon over a local Unix socket. - Graphical configuration editor covering general settings, Obsidian vault options, AI/Ollama settings, audio parameters, prompts and note templates. - Obsidian integration: callouts by note type, wikilink injection with case-insensitive word-boundary matching, tag generation, YAML frontmatter parsing and merging without duplication, and configurable folder organization (none, month, week, month+week). - Multilingual dictation with a configurable output language for titles, prose, wikilinks and tags. - Customizable Markdown prompts for each pipeline phase and retry handling, plus three note templates with placeholders for title, text, tags, date and time. - Privacy-oriented design: no temporary audio files are written to disk; audio stays in memory. - Retry and validation logic for structured JSON output from the model, with configurable retry count. - XDG-compliant rotating logs. Requirements and constraints: Ollama running locally with a pulled Gemma 4 model (default gemma4:e4b-it-qat), PortAudio installed, and a desktop notification service. The README states Gemma 4 audio models support clips up to 30 seconds, and the tool is optimized for dictations within that limit. Installation is offered via pipx or uv from GitHub, or as an editable local install for development.