fix: make heavy ML libraries import conditionally on Render to prevent OOM/timeouts

app/core/config.py

@@ -1,10 +1,16 @@
 import os
 from dotenv import load_dotenv
 from langchain_groq import ChatGroq
-from langchain_huggingface import HuggingFaceEmbeddings
 
 load_dotenv()
 
+IS_RENDER = os.getenv("RENDER") == "true"
+
+if not IS_RENDER:
+    from langchain_huggingface import HuggingFaceEmbeddings
+else:
+    from langchain_google_genai import GoogleGenAIEmbeddings
+
 BASE_DIR = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
 DATA_PATH = os.path.join(BASE_DIR, "data")
 QDRANT_PATH = os.path.join(BASE_DIR, "VectorDB")
@@ -20,6 +26,8 @@ LLM_MODEL_NAME = "llama-3.3-70b-versatile"
 RERANK_MODEL_NAME = "cross-encoder/ms-marco-MiniLM-L-6-v2"
 
 def get_embeddings():
+    if IS_RENDER:
+        return GoogleGenAIEmbeddings(model="models/text-embedding-004")
     return HuggingFaceEmbeddings(model_name=EMBED_MODEL_NAME)
 
 

app/service/retriever.py

@@ -1,7 +1,6 @@
 from qdrant_client import QdrantClient
 from qdrant_client.http import models
-from sentence_transformers import CrossEncoder
-from app.core.config import QDRANT_PATH, RERANK_MODEL_NAME, get_embeddings, COLLECTION_NAME
+from app.core.config import QDRANT_PATH, RERANK_MODEL_NAME, get_embeddings, COLLECTION_NAME, IS_RENDER
 
 _rerank_model = None
 _qdrant_client = None
@@ -9,9 +8,11 @@ _qdrant_client = None
 def get_rerank_model():
     global _rerank_model
     if _rerank_model is None:
+        from sentence_transformers import CrossEncoder
         _rerank_model = CrossEncoder(RERANK_MODEL_NAME, device='cpu')
     return _rerank_model
 
+
 def get_qdrant_client():
     global _qdrant_client
     if _qdrant_client is None:
@@ -59,6 +60,10 @@ def apply_post_retrieval_rerank(user_query: str, retrieved_texts: str, top_n: in
     if not doc_list:
         return "Không tìm thấy ngữ cảnh phù hợp."
 
+    if IS_RENDER:
+        # Bỏ qua CrossEncoder khi chạy trên Render để tránh tràn bộ nhớ (Free tier 512MB RAM)
+        return "\n\n---\n\n".join(doc_list[:top_n])
+
     rerank_model = get_rerank_model()
     pairs = [[user_query, doc] for doc in doc_list]
     scores = rerank_model.predict(pairs)