enable cuda and added pyinstrument

main.py

@@ -1,5 +1,7 @@
 import os
 import asyncio
+import torch
+
 from contextlib import asynccontextmanager
 
 from typing import AsyncGenerator, List, Optional, Tuple, Dict, Any
@@ -97,6 +99,7 @@ def convert_messages_langchain(messages: List[ChatMessage]):
     return list_chatmessages
 
 
+print(f"CUDA available: {torch.cuda.is_available()}")
 champ = ChampService(base_dir=BASE_DIR, hf_token=HF_TOKEN)
 
 

rag.py

@@ -0,0 +1,34 @@
+# app/champ/rag.py
+import torch
+
+from pathlib import Path
+
+from langchain_community.vectorstores import FAISS as LCFAISS
+from langchain_huggingface import HuggingFaceEmbeddings
+
+from constants import BASE_DIR, HF_TOKEN
+
+
+def load_vector_store(
+    base_dir: Path = BASE_DIR,
+    hf_token: str = HF_TOKEN,
+    rag_relpath: str = "rag_data/FAISS_ALLEN_20260129",
+    embedding_model: str = "BAAI/bge-large-en-v1.5",
+    device: str = "cuda" if torch.cuda.is_available() else "cpu",
+) -> LCFAISS:
+    rag_path = base_dir / rag_relpath
+
+    model_embedding_kwargs = {"device": device, "use_auth_token": hf_token}
+    encode_kwargs = {"normalize_embeddings": True}
+
+    embeddings = HuggingFaceEmbeddings(
+        model_name=embedding_model,
+        model_kwargs=model_embedding_kwargs,
+        encode_kwargs=encode_kwargs,
+    )
+
+    return LCFAISS.load_local(
+        str(rag_path),
+        embeddings,
+        allow_dangerous_deserialization=True,  # safe because you built the files
+    )

requirements.txt

@@ -125,3 +125,4 @@ xxhash==3.6.0
 yarl==1.22.0
 zstandard==0.25.0
 pytz==2025.2
+pyinstrument==5.1.2