From d234ef394e0e69b853641309e26bf8d1518de3d7 Mon Sep 17 00:00:00 2001 From: Jake Date: Fri, 1 May 2026 09:59:37 +0100 Subject: [PATCH] docs(readme): refresh bundled LLM list to Qwen3.5/3.6 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit README LLM-formatting section now states four Qwen tiers (Qwen3.5 2B / 4B / 9B + Qwen3.6 27B) and the magnotia-llm crate row reflects the four-tier registry. The whisper-ecosystem context doc gets the same refresh and cites unsloth as the GGUF source. Older roadmap and Phase-0 audit docs left untouched — they are dated historical artefacts and rewriting them would muddy the audit trail. Co-Authored-By: Claude Opus 4.7 (1M context) --- README.md | 4 ++-- docs/whisper-ecosystem/magnotia-context.md | 4 ++-- 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/README.md b/README.md index 54282d6..536e894 100644 --- a/README.md +++ b/README.md @@ -41,7 +41,7 @@ These are enforced in the codebase (where practical) and in the docs under [`doc ### LLM formatting (local only) - Local LLM runtime via [llama-cpp-2](https://github.com/utilityai/llama-cpp-rs) 0.1.144 with Vulkan. -- Three Qwen3 tiers (1.7B, 4B-Instruct-2507, 14B) auto-selected by hardware probe. +- Four Qwen tiers (Qwen3.5 2B / 4B / 9B + Qwen3.6 27B) auto-selected by hardware probe. - GBNF grammar-constrained output for task extraction (always-parseable JSON). - System prompt hardened against voice-delivered prompt injection. @@ -167,7 +167,7 @@ magnotia/ | **`magnotia-core`** | Shared types (`Segment`, `Transcript`, `Megabytes`, `ModelId`), constants, the `Engine` / `SpeedTier` / `AccuracyTier` enums, hardware probe (`sysinfo`-based), model registry (Whisper + Parakeet entries), hardware-aware recommendation scoring, `process_watch` for meeting detection. | | **`magnotia-audio`** | `cpal`-based microphone capture with device hotplug + error forwarding, VAD, `rubato` streaming resampler to 16 kHz mono, `symphonia` file decoding, `hound` WAV I/O. | | **`magnotia-transcription`** | `whisper-rs` backend (`WhisperRsBackend`) that owns a `WhisperContext` and supports `set_initial_prompt`. `LocalEngine` wraps both Whisper and Parakeet (via `transcribe-rs` ONNX) behind a common `Transcriber` trait. Streaming primitives (`VadChunker`, `LocalAgreement`, buffer trim) live in the `streaming/` module. Model manager handles downloads, paths, and disk checks. | -| **`magnotia-llm`** | `llama-cpp-2` engine with Qwen3 model manager. Three high-level surfaces: `cleanup_text` (formatting), `decompose_task` (3–7 micro-steps, GBNF-constrained JSON array), `extract_tasks` (optional-array, GBNF-constrained). Resumable HTTP downloads with SHA-256 verify. | +| **`magnotia-llm`** | `llama-cpp-2` engine with a four-tier Qwen3.5 / Qwen3.6 model manager. Three high-level surfaces: `cleanup_text` (formatting), `decompose_task` (3–7 micro-steps, GBNF-constrained JSON array), `extract_tasks` (optional-array, GBNF-constrained). Resumable HTTP downloads with SHA-256 verify. | | **`magnotia-ai-formatting`** | Post-processing pipeline: filler removal, British English conversion, anti-hallucination filter, smart paragraph breaks on long pauses, optional LLM cleanup. Also hosts the `llm_client::CLEANUP_PROMPT` constant (prompt-injection-hardened). | | **`magnotia-storage`** | SQLite via `sqlx` 0.8. Migrations, CRUD for transcripts / tasks / subtasks / profiles / profile terms / settings / error log, FTS5 search, file-storage paths. | | **`magnotia-hotkey`** | Linux `evdev` hotkey listener with device hotplug. Parses Tauri-style hotkey strings (`Ctrl+Shift+R`), emits Pressed / Released events. Works natively on Wayland (no X11 dependency). Checks `/dev/input/event*` access on startup; surfaces a clear "add yourself to the `input` group" error when missing. | diff --git a/docs/whisper-ecosystem/magnotia-context.md b/docs/whisper-ecosystem/magnotia-context.md index 13dc1c1..e71b7f8 100644 --- a/docs/whisper-ecosystem/magnotia-context.md +++ b/docs/whisper-ecosystem/magnotia-context.md @@ -78,9 +78,9 @@ Magnotia uses `whisper-rs` 0.16 with the `vulkan` feature ([crates/transcription The `brief.md` references faster-whisper as a pattern source — treat it as one. Do not adopt it as a dependency. -### LLM runtime: `llama-cpp-2` with Qwen3 tiers +### LLM runtime: `llama-cpp-2` with Qwen3.5 / Qwen3.6 tiers -`crates/llm/` owns the LLM runtime. Three tiers (Qwen3 1.7B, Qwen3 4B-Instruct-2507, Qwen3 14B) selected via hardware probe. Resumable HTTP downloads with SHA-256 verification. Do not add alternative runtimes (candle, mistral.rs, etc.) without strong evidence — the current stack is working. +`crates/llm/` owns the LLM runtime. Four tiers (Qwen3.5 2B, Qwen3.5 4B, Qwen3.5 9B, Qwen3.6 27B) selected via hardware probe. All Q4_K_M GGUFs from unsloth's HF org. Resumable HTTP downloads with SHA-256 verification. Do not add alternative runtimes (candle, mistral.rs, etc.) without strong evidence — the current stack is working. ### GPU compatibility: Vulkan everywhere