Deploy Small Talk: Reachy Mini AI podcast (gradio.Server + three.js + LiveKit)

backend/showgen.py

@@ -14,6 +14,7 @@ import re
 import tempfile
 
 import httpx
+from better_profanity import profanity
 
 from . import config
 
@@ -48,20 +49,15 @@ MODERATE_SYS = (
     '{"safe": true|false, "reason": "<short reason if unsafe>"}'
 )
 
-# cheap last line of defence so the obvious stuff dies even if the LLM is down
-_BLOCKLIST = re.compile(
-    r"\b(nsfw|porn\w*|hentai|xxx|blowjob|cumshot|deepthroat|gangbang|paedo\w*|pedo\w*|"
-    r"rape\w*|bestiality|loli\w*|nigger|faggot)\b", re.IGNORECASE)
-
-
 async def moderate_topic(title: str, topic: str) -> tuple[bool, str]:
     """One fast LLM pass: is this title/topic okay for a public, all-ages show?
 
-    Fails OPEN on LLM errors (the regex blocklist still applies) so a cold or
-    flaky Modal endpoint can't take room creation down with it.
+    Fails OPEN on LLM errors (the better-profanity wordlist still applies) so a
+    cold or flaky Modal endpoint can't take room creation down with it.
     """
     text = f"{title}\n{topic}".strip()
-    if _BLOCKLIST.search(text):
+    # wordlist first: the obvious stuff dies instantly, even with the LLM down
+    if profanity.contains_profanity(text):
         return False, "that topic isn't suitable for a public show"
     try:
         async with httpx.AsyncClient(timeout=90) as cx:

requirements.txt

@@ -4,3 +4,4 @@ livekit-api>=1.0
 numpy>=1.26
 python-dotenv>=1.0
 python-multipart>=0.0.9
+better-profanity>=0.7