Deploy SaulGPT backend with full ChromaDB, Supabase, and Gradio API server

app.py

@@ -66,6 +66,13 @@ try:
 except Exception as _e:
     print(f"init_db check: {_e}")
 
+# Pre-warm retrieval models on CPU
+try:
+    from layer2_retrieval import _ensure_models
+    _ensure_models()
+except Exception as _e:
+    print(f"model pre-warm check: {_e}")
+
 if __name__ == "__main__":
     # Launch Gradio server in non-blocking mode to attach FastAPI routes
     server_app, local_url, share_url = demo.launch(prevent_thread_lock=True)

backend/layer2_retrieval.py

@@ -91,14 +91,14 @@ def _ensure_models():
 
     try:
         from sentence_transformers import SentenceTransformer
-        embedder = SentenceTransformer(EMBEDDING_MODEL)
+        embedder = SentenceTransformer(EMBEDDING_MODEL, device="cpu")
     except Exception as e:
         print(f"[Layer 2 Warning] SentenceTransformer load failed: {e}")
         embedder = None
 
     try:
         from sentence_transformers import CrossEncoder
-        reranker = CrossEncoder(RERANKER_MODEL)
+        reranker = CrossEncoder(RERANKER_MODEL, device="cpu")
     except Exception as e:
         print(f"[Layer 2 Warning] CrossEncoder load skipped: {e}")
         reranker = None

backend/layer4_validation.py

@@ -144,7 +144,8 @@ def _ensure_nli():
             "text-classification",
             model=NLI_MODEL_NAME,
             tokenizer=NLI_MODEL_NAME,
-            top_k=None
+            top_k=None,
+            device="cpu"
         )
         nli_model_available = True
         print("[Layer 4] NLI model loaded.\n")