File size: 18,218 Bytes
8732fc1 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 | """Vibe Search: search your library by how music sounds and feels ("rainy Sunday coffee shop", "upbeat but
melancholic, no vocals"), or find songs with the same vibe as one you pick, and save the results as a playlist.
It runs next to AudioMuse-AI's own models and changes nothing in them:
* after each song is analysed, a small audio model (EfficientAT mn20_as, distilled from a teacher that joins
LCO-Embedding-Omni-7B and e5-omni-7B) gives it a 512-number vibe vector, stored in this plugin's own table;
* "Index my library" fills in the songs that were analysed before the plugin was installed;
* searches are turned into the same kind of vector by a text model (Qwen3-Embedding-0.6B plus a small trained layer)
and compared with every song.
The models are downloaded on first use from the Hugging Face repository set on the settings page.
"""
import html
import os
import threading
import time
import numpy as np
from flask import Blueprint, redirect, request, url_for
from plugin.api import (config, enqueue, get_db, get_score_data_by_ids, get_setting, logger, manage_plugins_url,
render_page, set_setting, table)
from . import vibemodel
bp = Blueprint("vibe_search", __name__)
DEFAULTS = {
"model_repo": "Divi19/vibe-embed",
"revision": "main",
"base_url": "https://huggingface.co",
"precision": "int8",
"threads": 4,
"embed_on_analysis": True,
"result_limit": 50,
}
STUDENT = "mn20_as" # the only audio model size the release ships
INT8_TEXT_MIN_ORT = (1, 28) # the 8-bit text model uses an operator setting that older onnxruntime builds lack
def setting(key):
return get_setting(key, DEFAULTS[key])
def ort_version():
try:
import onnxruntime as ort
return tuple(int(x) for x in ort.__version__.split(".")[:2])
except Exception:
return (0, 0)
def model_dir():
base = getattr(config, "PLUGINS_DIR", "") or "/tmp"
return os.path.join(base, "_data", "vibe_search", "models", setting("model_repo").replace("/", "__"), setting("revision"))
# ---------------------------------------------------------------- models (loaded once per container, on first use)
_models, _model_errors, _model_notes, _model_lock = {}, {}, {}, threading.Lock()
def get_model(kind):
"""kind: 'audio' (worker: indexing) or 'text' (web: searching). Returns the loaded model or None (error logged)."""
wanted = setting("precision")
key = (kind, setting("model_repo"), setting("revision"), STUDENT, wanted)
with _model_lock:
if key in _models:
return _models[key]
tries = [wanted] if wanted == "fp32" else ["int8", "fp32"]
if kind == "text" and wanted == "int8" and ort_version() < INT8_TEXT_MIN_ORT:
tries = ["fp32"] # e.g. AudioMuse-AI's no-AVX2 image (onnxruntime 1.19): skip the 600 MB download that cannot run
_model_notes[kind] = (f"The search model runs at full precision: this onnxruntime ({'.'.join(map(str, ort_version()))}) "
f"is older than {'.'.join(map(str, INT8_TEXT_MIN_ORT))}, which the 8-bit file needs.")
errors = []
for precision in tries:
try:
d = model_dir()
vibemodel.download_files(d, setting("base_url"), setting("model_repo"), setting("revision"),
vibemodel.files_needed(STUDENT, precision, [kind]), log=logger.info)
m = vibemodel.VibeModel(d, STUDENT, precision, setting("threads"))
cos = m.load_audio() if kind == "audio" else m.load_text()
logger.info(f"vibe_search: {kind} model ready ({m.tag}, {precision}, check {cos:.4f})")
if precision != wanted and kind not in _model_notes:
_model_notes[kind] = f"The {kind} model runs at full precision because the 8-bit file failed: {errors[-1]}"
_models[key] = m
_model_errors.pop(kind, None)
return m
except Exception as e:
errors.append(f"{precision}: {type(e).__name__}: {e}")
logger.exception(f"vibe_search: the {kind} model ({precision}) could not be loaded")
_model_errors[kind] = "; ".join(errors)
return None
def current_tag():
"""The model version stored vectors must match. Read from config.json without loading a model."""
try:
import json, hashlib
cfg = json.load(open(os.path.join(model_dir(), "config.json")))
return f"{STUDENT}:{hashlib.md5(json.dumps(cfg, sort_keys=True).encode()).hexdigest()[:12]}"
except Exception:
return None
# ---------------------------------------------------------------- storage
def vec_table():
return table("vectors")
def migrate(db):
cur = db.cursor()
cur.execute("CREATE TABLE IF NOT EXISTS " + vec_table() + " (item_id TEXT PRIMARY KEY, model_tag TEXT NOT NULL, "
"vec BYTEA NOT NULL, n_windows INTEGER, updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP)")
cur.execute("CREATE INDEX IF NOT EXISTS " + vec_table() + "_tag ON " + vec_table() + " (model_tag)")
db.commit()
cur.close()
def save_vector(item_id, tag, vec, n_windows):
db = get_db()
cur = db.cursor()
cur.execute("INSERT INTO " + vec_table() + " (item_id, model_tag, vec, n_windows) VALUES (%s, %s, %s, %s) "
"ON CONFLICT (item_id) DO UPDATE SET model_tag = EXCLUDED.model_tag, vec = EXCLUDED.vec, "
"n_windows = EXCLUDED.n_windows, updated_at = CURRENT_TIMESTAMP",
(str(item_id), tag, np.asarray(vec, np.float16).tobytes(), int(n_windows)))
db.commit()
cur.close()
_matrix_cache.clear()
def indexed_ids(tag):
cur = get_db().cursor()
cur.execute("SELECT item_id FROM " + vec_table() + " WHERE model_tag = %s", (tag,))
out = {r[0] for r in cur.fetchall()}
cur.close()
return out
_matrix_cache = {}
def song_matrix(tag):
"""All stored song vectors of this model version, cached until a vector is added."""
cur = get_db().cursor()
cur.execute("SELECT COUNT(*), MAX(updated_at) FROM " + vec_table() + " WHERE model_tag = %s", (tag,))
stamp = tuple(str(x) for x in cur.fetchone())
if _matrix_cache.get("stamp") == (tag,) + stamp:
cur.close()
return _matrix_cache["ids"], _matrix_cache["mat"]
cur.execute("SELECT item_id, vec FROM " + vec_table() + " WHERE model_tag = %s", (tag,))
rows = cur.fetchall()
cur.close()
ids = [r[0] for r in rows]
mat = np.stack([np.frombuffer(bytes(r[1]), np.float16) for r in rows]) if rows else np.zeros((0, 512), np.float16)
_matrix_cache.update(stamp=(tag,) + stamp, ids=ids, mat=mat)
return ids, mat
def top_k(query_vec, tag, k, exclude=None):
ids, mat = song_matrix(tag)
if not ids:
return []
scores = np.concatenate([mat[i:i + 50000].astype(np.float32) @ query_vec for i in range(0, len(ids), 50000)])
order = np.argsort(-scores)
out = []
for j in order:
if ids[j] != exclude:
out.append((ids[j], float(scores[j])))
if len(out) >= k:
break
return out
# ---------------------------------------------------------------- audio
def load_wave(path):
"""Decode to 32 kHz mono with AudioMuse-AI's own loader (the same one its analysis uses)."""
try:
from tasks.analysis.song import robust_load_audio_with_fallback
audio, sr = robust_load_audio_with_fallback(path, target_sr=vibemodel.SR)
if audio is not None and sr == vibemodel.SR:
return np.asarray(audio, np.float32)
except Exception:
logger.exception("vibe_search: AudioMuse-AI's audio loader failed; trying soundfile")
import soundfile as sf
y, sr = sf.read(path, dtype="float32", always_2d=True)
y = y.mean(1)
if sr != vibemodel.SR:
import soxr
y = soxr.resample(y, sr, vibemodel.SR, quality="HQ").astype(np.float32)
return y
def canonical_id(item_id, server_id=None):
try:
from tasks.mediaserver.registry import canonical_input_ids
return canonical_input_ids([item_id], server_id).get(str(item_id), str(item_id))
except Exception:
return str(item_id)
def embed_file(path, item_id):
m = get_model("audio")
if m is None:
return False
t0 = time.time()
vec, n = m.embed_audio(load_wave(path))
save_vector(item_id, m.tag, vec, n)
logger.info(f"vibe_search: {item_id} embedded ({n} windows, {time.time() - t0:.1f} s)")
return True
def on_song_analyzed(song):
"""Runs on the worker right after AudioMuse-AI finishes a song; the audio file is deleted once this returns."""
if not setting("embed_on_analysis") or not song.get("audio_path"):
return
embed_file(song["audio_path"], canonical_id(song["item_id"], song.get("server_id")))
def index_library():
"""Worker task: give a vibe vector to every song that does not have one yet (for songs analysed before install)."""
from tasks.mediaserver import download_track, get_all_songs
m = get_model("audio")
if m is None:
raise RuntimeError("the audio model could not be loaded: " + _model_errors.get("audio", "unknown error"))
songs = get_all_songs() or []
done = indexed_ids(m.tag)
todo = [s for s in songs if canonical_id(s.get("Id")) not in done]
status = {"started": time.strftime("%Y-%m-%d %H:%M"), "total": len(todo), "done": 0, "failed": 0, "finished": False}
set_setting("index_status", status)
logger.info(f"vibe_search: indexing {len(todo)} of {len(songs)} songs")
temp_dir = getattr(config, "TEMP_DIR", "/tmp")
os.makedirs(temp_dir, exist_ok=True)
for i, item in enumerate(todo):
path = None
try:
path = download_track(temp_dir, item)
if not path or not embed_file(path, canonical_id(item.get("Id"))):
status["failed"] += 1
else:
status["done"] += 1
except Exception:
status["failed"] += 1
logger.exception(f"vibe_search: could not index {item.get('Id')}")
finally:
if path and os.path.exists(path):
os.remove(path)
if (i + 1) % 25 == 0:
set_setting("index_status", status)
status["finished"] = True
set_setting("index_status", status)
# ---------------------------------------------------------------- pages
CSS = """<style>
.vs { max-width: 980px; display: grid; gap: 16px; }
.vs form.search { display: flex; gap: 8px; flex-wrap: wrap; }
.vs input[type=text] { flex: 1 1 360px; padding: 8px 10px; font-size: 1rem; }
.vs table { border-collapse: collapse; width: 100%; }
.vs td, .vs th { padding: 6px 8px; border-bottom: 1px solid rgba(128,128,128,.25); text-align: left; vertical-align: top; }
.vs .score { font-variant-numeric: tabular-nums; opacity: .7; }
.vs .note { opacity: .75; font-size: .92rem; }
.vs .err { color: #b3261e; }
</style>"""
def page(body, title="Vibe Search"):
return render_page(CSS + '<div class="vs">' + body + "</div>", title=title)
def status_html():
tag = current_tag()
n_songs = n_vec = 0
try:
cur = get_db().cursor()
cur.execute("SELECT COUNT(*) FROM score")
n_songs = cur.fetchone()[0]
if tag:
cur.execute("SELECT COUNT(*) FROM " + vec_table() + " WHERE model_tag = %s", (tag,))
n_vec = cur.fetchone()[0]
cur.close()
except Exception:
logger.exception("vibe_search: status query failed")
st = get_setting("index_status", None)
msg = f"{n_vec:,} of {n_songs:,} analysed songs have a vibe vector."
if st and not st.get("finished"):
msg += f" Indexing: {st['done']:,} of {st['total']:,} done" + (f", {st['failed']} failed" if st["failed"] else "") + "."
err = "".join(f'<p class="err">The {k} model is not available: {html.escape(v)}</p>' for k, v in _model_errors.items())
notes = "".join(f'<p class="note">{html.escape(v)}</p>' for v in _model_notes.values())
return f'<p class="note">{msg}</p>{notes}{err}'
def results_html(rows, heading, playlist_name):
if not rows:
return "<p>No songs with a vibe vector yet. Analyse your library, or use Index my library below.</p>"
info = {r["item_id"]: r for r in (get_score_data_by_ids([i for i, _ in rows]) or [])}
trs = []
for item_id, score in rows:
r = info.get(item_id, {})
trs.append(
f'<tr><td><input type="checkbox" name="ids" value="{html.escape(item_id)}" checked></td>'
f'<td>{html.escape(str(r.get("title") or item_id))}</td><td>{html.escape(str(r.get("author") or ""))}</td>'
f'<td>{html.escape(str(r.get("album") or ""))}</td><td class="score">{score:.3f}</td>'
f'<td><a href="{url_for("vibe_search.similar", item_id=item_id)}">same vibe</a></td></tr>')
return (f"<h3>{html.escape(heading)}</h3>"
f'<form method="post" action="{url_for("vibe_search.playlist")}">'
f"<table><tr><th></th><th>Title</th><th>Artist</th><th>Album</th><th>Match</th><th></th></tr>{''.join(trs)}</table>"
f'<p><input type="text" name="name" value="{html.escape(playlist_name)}" aria-label="Playlist name"> '
f'<button type="submit">Save ticked songs as a playlist</button></p></form>')
def search_form(q=""):
return (f'<form class="search" method="get" action="{url_for("vibe_search.home")}">'
f'<input type="text" name="q" value="{html.escape(q)}" placeholder="late night drive, rainy Sunday, upbeat but melancholic..." '
f'aria-label="Describe the music"><button type="submit">Search</button></form>')
@bp.route("/")
def home():
q = (request.args.get("q") or "").strip()
body = "<h2>Vibe Search</h2>" + search_form(q) + status_html()
if q:
m = get_model("text")
if m is None:
body += '<p class="err">The search model could not be loaded; see the message above.</p>'
else:
tag = current_tag() or m.tag
rows = top_k(m.embed_texts([q])[0], tag, int(setting("result_limit")))
body += results_html(rows, f"Songs that match “{q}”", f"Vibe: {q[:60]}")
body += (f'<form method="post" action="{url_for("vibe_search.index")}"><button type="submit">Index my library</button> '
'<span class="note">Gives a vibe vector to analysed songs that do not have one yet (runs in the background).</span></form>')
return page(body)
@bp.route("/similar/<path:item_id>")
def similar(item_id):
tag = current_tag()
ids, mat = song_matrix(tag) if tag else ([], None)
body = "<h2>Vibe Search</h2>" + search_form() + status_html()
if item_id not in ids:
body += "<p>This song has no vibe vector yet.</p>"
else:
seed = mat[ids.index(item_id)].astype(np.float32)
r = (get_score_data_by_ids([item_id]) or [{}])[0]
name = f'{r.get("title") or item_id}' + (f' by {r.get("author")}' if r.get("author") else "")
body += results_html(top_k(seed, tag, int(setting("result_limit")), exclude=item_id),
f"Same vibe as {name}", f"Same vibe: {name[:60]}")
return page(body)
@bp.route("/playlist", methods=["POST"])
def playlist():
ids = request.form.getlist("ids")
name = (request.form.get("name") or "Vibe Search").strip()[:120]
if not ids:
return page("<h2>Vibe Search</h2><p>No songs were ticked.</p>" + search_form())
try:
from tasks.mediaserver import create_or_replace_playlist
create_or_replace_playlist(name, ids)
msg = f"<p>Saved {len(ids)} songs as the playlist “{html.escape(name)}” on your media server.</p>"
except Exception as e:
logger.exception("vibe_search: playlist creation failed")
msg = f'<p class="err">The playlist could not be created: {html.escape(str(e))}</p>'
return page("<h2>Vibe Search</h2>" + msg + search_form())
@bp.route("/index", methods=["POST"])
def index():
try:
enqueue(index_library)
msg = "<p>Indexing started in the background. This page shows the progress.</p>"
except Exception as e:
msg = f'<p class="err">Indexing could not start now: {html.escape(str(e))}</p>'
return page("<h2>Vibe Search</h2>" + msg + search_form() + status_html())
@bp.route("/settings", methods=["GET", "POST"])
def settings():
if request.method == "POST":
for k, v in DEFAULTS.items():
if k not in request.form and not isinstance(v, bool):
continue
raw = request.form.get(k)
val = (raw == "on") if isinstance(v, bool) else (int(raw) if isinstance(v, int) else raw.strip())
set_setting(k, val)
_models.clear(); _model_errors.clear(); _model_notes.clear(); _matrix_cache.clear()
return redirect(manage_plugins_url())
rows = []
for k, v in DEFAULTS.items():
cur = setting(k)
if isinstance(v, bool):
field = f'<input type="checkbox" name="{k}" id="{k}"{" checked" if cur else ""}>'
else:
field = f'<input type="text" name="{k}" id="{k}" value="{html.escape(str(cur))}">'
rows.append(f'<tr><td><label for="{k}">{k}</label></td><td>{field}</td></tr>')
note = ("<p class=\"note\">precision: int8 (smaller, faster) or fp32. With int8, a model that cannot run on this "
"installation (for example the 8-bit search model on onnxruntime older than 1.28) falls back to fp32 by itself. "
"Changing the model repository or revision makes stored vectors outdated: run "
"Index my library afterwards. Models are stored in " + html.escape(model_dir()) + ".</p>")
return page(f'<h2>Vibe Search settings</h2><form method="post"><table>{"".join(rows)}</table>'
f'<button type="submit">Save</button></form>{note}', title="Vibe Search settings")
def register(ctx):
ctx.on_install(migrate)
ctx.add_blueprint(bp)
ctx.add_menu_item("Vibe Search", "vibe_search.home")
ctx.on_song_analyzed(on_song_analyzed)
ctx.add_task("index_library", index_library)
|