Enable direct CDN downloads of quantized MiniLM vectors

backend/app.py

@@ -131,8 +131,8 @@ def health():
 
 
 _ASSET_PATH = re.compile(
-    r"(?:openai/)?(?:manifest\.json|(?:ICD10CM|ICD9CM|ICPC|MDR|RCD2|SNOMEDCT_US)-[0-9a-f]{20}/"
-    r"(?:lexical\.json\.gz|[0-9]{7}\.(?:json\.gz|f32)))\Z"
+    r"(?:(?:openai|minilm-int8)/)?(?:manifest\.json|(?:ICD10CM|ICD9CM|ICPC|MDR|RCD2|SNOMEDCT_US)-[0-9a-f]{20}/"
+    r"(?:lexical\.json\.gz|[0-9]{7}\.(?:json\.gz|f32|i8)))\Z"
 )
 
 
@@ -155,7 +155,7 @@ def terminology_asset(asset_path: str):
     client = httpx.Client(timeout=60, follow_redirects=True)
     url = f"https://huggingface.co/datasets/{repo}/resolve/{revision}/{asset_path}"
     public = os.getenv("PIPELINE_PUBLIC_ACCESS", "").lower() == "true"
-    if public and asset_path.endswith(".f32"):
+    if public and asset_path.endswith((".f32", ".i8")):
         # Let the CDN serve large vectors directly. Only the temporary link to
         # this file reaches the browser; the HF access token stays server-side.
         try: