|
Download plugin/README.md from Divi19/vibe-embed: direct link, hf CLI and curl.
- Browser
- Download file 4.43 kB
-
https://huggingface.co/Divi19/vibe-embed/resolve/main/plugin/README.md
- Command line
-
hf download hf://Divi19/vibe-embed/plugin/README.md
-
curl -L -o README.md https://huggingface.co/Divi19/vibe-embed/resolve/main/plugin/README.md
4.43 kB
| # Vibe Search for AudioMuse-AI | |
| Search your own music library by how songs sound and feel ("rainy Sunday coffee shop", "upbeat but melancholic, no vocals", "something for a long night drive"). You can also find songs with the same vibe as one you pick, and save any result list as a playlist on your media server. | |
| The model is small and runs on the CPU. It sits next to AudioMuse-AI's own analysis and leaves it alone: the plugin keeps its vectors in its own table and never edits AudioMuse-AI's data. | |
| ## What it does | |
| - **Vibe vectors for every song.** Each time AudioMuse-AI analyses a song, the plugin runs a small audio model (EfficientAT mn20_as, distilled from a combined LCO-Embedding-Omni-7B + e5-omni-7B teacher) over the song in 10-second windows. It stores one 512-number vector per song. | |
| - **Index my library.** This fills in vectors for songs that were analysed before the plugin was installed. It runs as a background task and reports its progress. | |
| - **Search.** Your search text goes through the matching text model (Qwen3-Embedding-0.6B plus a small trained layer) and is compared with every song's vector. | |
| - **Same vibe.** Pick a song to list the songs whose vectors are closest to it. | |
| - **Playlists.** Save a result list to your media server in one click. | |
| Model version: v1.2. Benchmarks and limits are on the model card: https://huggingface.co/Divi19/vibe-embed | |
| ## Install | |
| Requires AudioMuse-AI 3.7.1 or later. | |
| 1. In AudioMuse-AI, open **Plugins**, add this repository URL, and refresh the catalog: | |
| `https://huggingface.co/Divi19/vibe-embed/resolve/main/plugin/manifest.json` | |
| 2. Install **Vibe Search**, then apply. AudioMuse-AI restarts its workers. | |
| 3. Open the Vibe Search page and press **Index my library** once. From then on, new songs get their vectors automatically when they are analysed. | |
| On first use, the plugin downloads its models into `PLUGINS_DIR/_data/vibe_search/models/`: about 35 MB for the audio model and about 818 MB for the 8-bit search model. It needs no extra Python packages: it uses the ONNX Runtime, numpy and tokenizers that AudioMuse-AI already ships with. | |
| ### ONNX Runtime versions | |
| - The 8-bit search model needs onnxruntime 1.28 or later, which the standard AudioMuse-AI images have. | |
| - On older onnxruntime builds (the no-AVX2 image ships 1.19), the plugin uses the full-precision search model instead (about 2396 MB, more RAM). It does this by itself and says so on its page. Results are the same. | |
| - If an 8-bit file fails to load for any other reason, the plugin falls back to full precision the same way. | |
| ## Settings | |
| | setting | default | meaning | | |
| |---|---|---| | |
| | model_repo | `Divi19/vibe-embed` | Hugging Face repository with the model files | | |
| | revision | `main` | branch, tag or commit of that repository | | |
| | base_url | `https://huggingface.co` | change this only for a mirror | | |
| | precision | `int8` | `int8` (smaller, faster) or `fp32` | | |
| | threads | 4 | CPU threads for the models | | |
| | embed_on_analysis | on | embed each song right after AudioMuse-AI analyses it | | |
| | result_limit | 50 | songs shown per search | | |
| If you change model_repo or revision, press **Index my library** again. Vectors from a different model are not mixed in: every vector is stored with the tag of the model that made it. Changing precision does not need re-indexing. | |
| ## Safety checks | |
| The model repository includes `parity.json`, which holds reference outputs recorded when the models were exported. Before first use, the plugin runs a test signal and five test searches through the downloaded models and compares the results (and the search tokens) with those references. If anything does not match (a wrong file, a broken download or a tokenizer mismatch), it refuses to run and shows the reason on its page. It does not silently return bad results. | |
| ## Model files | |
| The plugin expects this layout in the Hugging Face repository: | |
| ``` | |
| config.json | |
| parity.json | |
| mn20_as/audio_encoder_int8.onnx mn20_as/audio_encoder.onnx | |
| text/text_encoder_int8.onnx(.data) text/text_encoder.onnx(.data) | |
| text/tokenizer/tokenizer.json text/tokenizer/tokenizer_config.json | |
| ``` | |
| ## Licences | |
| - The plugin code is licensed under AGPL-3.0 (see LICENSE), the same as AudioMuse-AI, because it runs inside AudioMuse-AI. | |
| - The models it downloads are released separately under Apache-2.0. | |
| - The training music is CC0 / CC-BY / public-domain audio; the attribution list is in the model repository. | |