Merge pull request #166 from praekeltfoundation/put-embedding-model-b…

…ehind-flag

Put embedding model behind flag

contentrepo/settings/base.py

@@ -264,3 +264,5 @@
 EMAIL_SSL_CERTFILE = env.str("EMAIL_SSL_CERTFILE", None)
 EMAIL_SSL_KEYFILE = env.str("EMAIL_SSL_KEYFILE", None)
 EMAIL_TIMEOUT = env.int("EMAIL_TIMEOUT", None)
+
+LOAD_TRANSFORMER_MODEL = env.bool("LOAD_TRANSFORMER_MODEL", False)

home/constants.py

@@ -1,4 +1,5 @@
 # Define constants for use throughout the application
+from django.conf import settings
 from sentence_transformers import SentenceTransformer
 
 GENDER_CHOICES = [
@@ -21,4 +22,6 @@
     ("empty", "Empty"),
 ]
 
-model = SentenceTransformer("all-mpnet-base-v2")
+model = None
+if settings.LOAD_TRANSFORMER_MODEL:
+    model = SentenceTransformer("all-mpnet-base-v2")

home/models.py

@@ -620,6 +620,9 @@ def save_revision(
 def update_embedding(sender, instance, *args, **kwargs):
     from .utils import preprocess_content_for_embedding
 
+    if not model:
+        return
+
     embedding = {}
     if instance.enable_web:
         content = []

home/utils.py

@@ -82,6 +82,9 @@ def cosine_similarity(A, B):
 def retrieve_top_n_content_pieces(
     user_input, queryset, n=5, content_type=None, platform="web"
 ):
+    if not model:
+        return []
+
     # similar_embeddings = [{'faq_name':, 'faq_content':, 'embedding':}, ...] # We need to filter by content type and then retrieve their embeddings
     # Generate embedding for user text
     user_embedding = model.encode([user_input])