fix: set HF_HUB_OFFLINE=1 at runtime to prevent model network calls on startup

Dockerfile

@@ -23,9 +23,9 @@ COPY . .
 
 # Pin HuggingFace cache inside /app so the sentence-transformer model is
 # downloaded once during `docker build` and baked into the image layer.
-# This eliminates cold-start latency on HuggingFace Spaces restarts.
 ENV HF_HOME=/app/.hf_cache
 ENV TRANSFORMERS_CACHE=/app/.hf_cache/transformers
+ENV SENTENCE_TRANSFORMERS_HOME=/app/.hf_cache/sentence_transformers
 
 # Set ownership (includes .hf_cache written by the build step below)
 RUN chown -R appuser:appuser /app
@@ -35,6 +35,13 @@ USER appuser
 # Pre-build FAISS index + download the embedding model into /app/.hf_cache
 RUN python src/build_faiss.py
 
+# ── Offline mode ────────────────────────────────────────────────────────────
+# Model is now cached in the image. Tell all HF libraries to NEVER call the
+# network at runtime — prevents "Could not resolve host: huggingface.co" errors.
+ENV TRANSFORMERS_OFFLINE=1
+ENV HF_DATASETS_OFFLINE=1
+ENV HF_HUB_OFFLINE=1
+
 # Environment defaults (override via HF Space secrets)
 ENV GROQ_API_KEY_1=""
 ENV GROQ_API_KEY_2=""

src/embeddings.py

@@ -6,8 +6,13 @@ from numpy.linalg import norm
 class EmbeddingsManager:
     # Khởi tạo model embedding từ cofig ngay khi gọi class
     def __init__(self):
+        import os
+        # Use cached model; never call the network (model is baked into Docker image)
+        local_only = os.getenv("TRANSFORMERS_OFFLINE", "0") == "1" or \
+                     os.getenv("HF_HUB_OFFLINE", "0") == "1"
         self.embeddings = HuggingFaceEmbeddings(
-            model_name=Config.EMBEDDING_MODEL
+            model_name=Config.EMBEDDING_MODEL,
+            model_kwargs={"local_files_only": local_only},
         )
     
     def get_embeddings(self):