Malina Claude Opus 4.7 commited on
Commit
87577ff
Β·
1 Parent(s): 3ffb602

NINA: privacy/cleanup for uploaded media (delete_cache + idle eviction)

Browse files

- Interface delete_cache=(300,300): Gradio purges uploaded videos older than
5 min, every 5 min. One upload still serves many prompts (audio + transcript
are cached in RAM, so the file isn't needed after first processing).
- In-memory session state evicted after 20 min idle; on-disk transcript/embedding
cache purged after 30 min. Keeps data from lingering; keeps multi-prompt fast.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Files changed (1) hide show
  1. app.py +33 -1
app.py CHANGED
@@ -68,8 +68,10 @@ import re
68
  import json
69
  import math
70
  import hashlib
 
71
  import tempfile
72
  import threading
 
73
  from concurrent.futures import ThreadPoolExecutor
74
 
75
  import ffmpeg
@@ -256,7 +258,31 @@ print(f"[NINA] Ready. ( {WHISPER_PRECISE} + CLIP warming up...)")
256
  # Stores segments, chunks (with embeddings), visual frame embeddings, the decoded
257
  # waveform, and a per-sector precise-transcription cache, keyed by video path.
258
  # Second prompt on same video skips transcription, chunking and CLIP entirely.
259
- _cache: dict = {} # { video_path: {"segments","chunks","visual","audio","precise"} }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
260
 
261
 
262
  # ── Audio ──────────────────────────────────────────────────────────────────────
@@ -1165,6 +1191,7 @@ def run(video_path: str, prompt: str):
1165
  prompt = prompt.strip()
1166
  max_words = _prompt_to_words(prompt) # output size inferred from the prompt
1167
  print(f'\n[NINA] ── New request ── "{prompt}" (target β‰ˆ {max_words} words)')
 
1168
 
1169
  # Stage 1–3: transcription + chunking + CLIP visual (cached per video)
1170
  fp = None
@@ -1193,6 +1220,7 @@ def run(video_path: str, prompt: str):
1193
  _disk_save(fp, entry)
1194
  print(f"[NINA] Cached transcript + visuals for this video.")
1195
 
 
1196
  segments = entry["segments"]
1197
  chunks = entry["chunks"]
1198
  visual = entry["visual"]
@@ -1295,6 +1323,10 @@ demo = gr.Interface(
1295
  "Answer length adapts to your prompt."
1296
  ),
1297
  flagging_mode="never",
 
 
 
 
1298
  )
1299
 
1300
  if __name__ == "__main__":
 
68
  import json
69
  import math
70
  import hashlib
71
+ import shutil
72
  import tempfile
73
  import threading
74
+ import time
75
  from concurrent.futures import ThreadPoolExecutor
76
 
77
  import ffmpeg
 
258
  # Stores segments, chunks (with embeddings), visual frame embeddings, the decoded
259
  # waveform, and a per-sector precise-transcription cache, keyed by video path.
260
  # Second prompt on same video skips transcription, chunking and CLIP entirely.
261
+ _cache: dict = {} # { video_path: {"segments","chunks","visual","audio","precise","_atime"} }
262
+
263
+ # Data lifecycle / privacy:
264
+ # β€’ The uploaded video itself is purged by Gradio's own temp cleaner
265
+ # (Interface(delete_cache=...)) β€” see the bottom of this file β€” so a file is
266
+ # removed shortly after the user's session goes idle. One upload still serves
267
+ # MANY prompts: the decoded audio + transcript live in RAM (below), so the
268
+ # file is no longer needed once processed.
269
+ # β€’ This in-memory state and the derived on-disk cache are TTL-evicted here so
270
+ # nothing lingers indefinitely.
271
+ CACHE_IDLE_TTL = 1200 # drop in-memory video state idle for > 20 min
272
+ CACHE_DISK_TTL = 1800 # delete on-disk transcript/embedding cache > 30 min old
273
+
274
+ def _evict_stale() -> None:
275
+ now = time.time()
276
+ for vp in [k for k, e in _cache.items() if now - e.get("_atime", now) > CACHE_IDLE_TTL]:
277
+ _cache.pop(vp, None)
278
+ try:
279
+ if os.path.isdir(CACHE_DIR):
280
+ for name in os.listdir(CACHE_DIR):
281
+ d = os.path.join(CACHE_DIR, name)
282
+ if os.path.isdir(d) and now - os.path.getmtime(d) > CACHE_DISK_TTL:
283
+ shutil.rmtree(d, ignore_errors=True)
284
+ except Exception:
285
+ pass
286
 
287
 
288
  # ── Audio ──────────────────────────────────────────────────────────────────────
 
1191
  prompt = prompt.strip()
1192
  max_words = _prompt_to_words(prompt) # output size inferred from the prompt
1193
  print(f'\n[NINA] ── New request ── "{prompt}" (target β‰ˆ {max_words} words)')
1194
+ _evict_stale() # drop idle sessions / stale disk cache
1195
 
1196
  # Stage 1–3: transcription + chunking + CLIP visual (cached per video)
1197
  fp = None
 
1220
  _disk_save(fp, entry)
1221
  print(f"[NINA] Cached transcript + visuals for this video.")
1222
 
1223
+ entry["_atime"] = time.time() # mark recently used (keeps many-prompt sessions alive)
1224
  segments = entry["segments"]
1225
  chunks = entry["chunks"]
1226
  visual = entry["visual"]
 
1323
  "Answer length adapts to your prompt."
1324
  ),
1325
  flagging_mode="never",
1326
+ # Privacy: Gradio purges uploaded files older than 5 min, every 5 min. One
1327
+ # upload still serves many prompts (audio/transcript are cached in RAM), and
1328
+ # the file is cleaned shortly after the session goes idle.
1329
+ delete_cache=(300, 300),
1330
  )
1331
 
1332
  if __name__ == "__main__":