refac: lazily load sentence_transformers to reduce start up memory usage

backend/apps/rag/main.py

@@ -48,8 +48,6 @@ import mimetypes
 import uuid
 import json
 
-import sentence_transformers
-
 from apps.webui.models.documents import (
     Documents,
     DocumentForm,
@@ -190,6 +188,8 @@ def update_embedding_model(
     update_model: bool = False,
 ):
     if embedding_model and app.state.config.RAG_EMBEDDING_ENGINE == "":
+        import sentence_transformers
+
         app.state.sentence_transformer_ef = sentence_transformers.SentenceTransformer(
             get_model_path(embedding_model, update_model),
             device=DEVICE_TYPE,
@@ -204,6 +204,8 @@ def update_reranking_model(
     update_model: bool = False,
 ):
     if reranking_model:
+        import sentence_transformers
+
         app.state.sentence_transformer_rf = sentence_transformers.CrossEncoder(
             get_model_path(reranking_model, update_model),
             device=DEVICE_TYPE,

backend/apps/rag/utils.py

@@ -442,8 +442,6 @@ from langchain_core.documents import BaseDocumentCompressor, Document
 from langchain_core.callbacks import Callbacks
 from langchain_core.pydantic_v1 import Extra
 
-from sentence_transformers import util
-
 
 class RerankCompressor(BaseDocumentCompressor):
     embedding_function: Any
@@ -468,6 +466,8 @@ class RerankCompressor(BaseDocumentCompressor):
                 [(query, doc.page_content) for doc in documents]
             )
         else:
+            from sentence_transformers import util
+
             query_embedding = self.embedding_function(query)
             document_embedding = self.embedding_function(
                 [doc.page_content for doc in documents]