Upload app.py

app.py

@@ -12,7 +12,6 @@ import time
 from pathlib import Path
 
 import gradio as gr
-import spaces  # Must import before configuring CUDA on ZeroGPU
 from google import genai
 from google.genai import types
 from pydantic import BaseModel, ConfigDict, Field, ValidationError
@@ -25,9 +24,9 @@ ROOT = Path(__file__).resolve().parent
 DATA = ROOT / 'assets/projects_clean.csv'
 VECTORS = ROOT / 'assets/project_vectors.npz'
 
-# On ZeroGPU, CUDA placement at module scope is supported by simulated CUDA.
-# Do not call model.encode() at startup or move the model to GPU inside callback.
-model = SentenceTransformer(MODEL_ID, device='cuda')
+# CPU retrieval avoids unauthenticated ZeroGPU per-user GPU quota for website visitors.
+# The frozen 103-project index, embedding model, ranking and /search contract are unchanged.
+model = SentenceTransformer(MODEL_ID, device='cpu')
 retriever = OriginRetrieval(DATA, VECTORS)
 
 class Idea(BaseModel):
@@ -91,13 +90,25 @@ def both(value):
     return {'en':value, 'ar':value}
 
 
-@spaces.GPU(duration=45)
 def search(title: str, description: str, language: str):
-    """GPU portion only. Returns a JSON-serializable dict (no Gemini call)."""
-    idea, lang = validate_idea(title, description, language)
-    vector = retriever.query_vector(model, idea.title, idea.description)
-    top = retriever.search_from_vector(vector, idea.title, idea.description, lang, top_k=5)
-    return retriever.public_retrieval(top, idea.title, idea.description, lang)
+    """CPU search. Uses the same embedding model/index; no ZeroGPU quota required."""
+    started = time.perf_counter()
+    stage = 'validation'
+    try:
+        idea, lang = validate_idea(title, description, language)
+        stage = 'embedding'
+        vector = retriever.query_vector(model, idea.title, idea.description)
+        stage = 'ranking'
+        top = retriever.search_from_vector(vector, idea.title, idea.description, lang, top_k=5)
+        result = retriever.public_retrieval(top, idea.title, idea.description, lang)
+        logger.info('origin.search_ok mode=cpu language=%s count=%d elapsed_ms=%d',
+                    lang, len(top), int((time.perf_counter()-started)*1000))
+        return result
+    except Exception as exc:
+        # Never log a student's title/description, credentials, or prompt.
+        logger.error('origin.search_failed mode=cpu stage=%s type=%s',
+                     stage, type(exc).__name__)
+        raise
 
 
 def enhance(title: str, description: str, language: str, search_payload: dict):
@@ -245,7 +256,7 @@ with gr.Blocks(title='Origin — ZeroGPU research demo') as demo:
     submit = gr.Button('Search and enhance', variant='primary')
     search_data = gr.JSON(label='Retrieved projects')
     final_data = gr.JSON(label='Idea Intelligence Report')
-    # Separate named Gradio endpoints, so Gemini runs AFTER GPU is released.
+    # Separate named endpoints. Gemini runs only after retrieval completes.
     submit.click(fn=search, inputs=[title,description,language], outputs=[search_data],
                  api_name='search', concurrency_limit=1).then(
                  fn=enhance, inputs=[title,description,language,search_data],