rollback: can't solve performance issues, rollback to lfc

app.py

@@ -8,18 +8,13 @@ from sentence_transformers import SentenceTransformer
 import gradio as gr
 from threading import Thread
 
-# FAISS Index laden (vorerst auf CPU)
-print("Loading FAISS index...")
-cpu_index = faiss.read_index("vector_db/index.faiss")
-gpu_index = None  # GPU Index wird später erstellt
-
+index = faiss.read_index("vector_db/index.faiss")
 with open("vector_db/chunks.pkl", "rb") as f:
     chunks = pickle.load(f)
 with open("vector_db/metadata.pkl", "rb") as f:
     metadata_dict = pickle.load(f)
 
 ST = SentenceTransformer("BAAI/bge-large-en-v1.5")
-
 github_base_url = "https://github.com/arsiba/EDPB-AI/blob/main/" 
 
 model_id = "HuggingFaceH4/zephyr-7b-beta"
@@ -37,7 +32,7 @@ model = AutoModelForCausalLM.from_pretrained(
     device_map={"": 0},
     torch_dtype=torch.bfloat16,
     trust_remote_code=True
-) 
+)
 
 SYS = (
     "You are a legal AI assistant specialized in GDPR/EDPB."
@@ -47,21 +42,8 @@ SYS = (
 
 @spaces.GPU()
 def retrieve(q, k=3):
-    global gpu_index, cpu_index
-    
-    # GPU Index beim ersten Aufruf erstellen
-    if gpu_index is None:
-        try:
-            print("Creating GPU index...")
-            res = faiss.StandardGpuResources()
-            gpu_index = faiss.index_cpu_to_gpu(res, 0, cpu_index)
-            print("FAISS index successfully moved to GPU")
-        except Exception as e:
-            print(f"Failed to move FAISS to GPU: {e}")
-            gpu_index = cpu_index  # Fallback to CPU
-    
     emb = ST.encode(q)
-    D, I = gpu_index.search(np.array([emb], dtype="float32"), k)
+    D, I = index.search(np.array([emb], dtype="float32"), k)
     docs, file_sources = [], []
     for i in I[0]:
         chunk = chunks[i]
@@ -70,6 +52,7 @@ def retrieve(q, k=3):
         file_sources.append(meta)
     return docs, file_sources
 
+
 def make_prompt(q, docs):
     context = "\n\n".join(f"Title: {d['title']}\nPages: {d['pages']}" for d in docs)
     prompt = f"detailed thinking off\n"
@@ -84,6 +67,7 @@ def build_markdown_links(file_input):
         lines.append(line)
     return "\n\n".join(lines)
 
+
 def build_markdown_chunks(docs):
     lines = []
     for idx, d in enumerate(docs, start=1):
@@ -118,7 +102,6 @@ def qa_fn(faiss_search, question, top_k, temperature, max_tokens):
         output = output.split("Output:", 1)[1].strip()
     return "\n# Generated Answer\n", output,"\n# Used Documents\n",  file_links, "\n# Used Context\n", markdown_chunks
 
-# Gradio Interface
 heading_answer  = gr.Markdown(label="Answer Heading")
 outputs_answer = gr.Textbox(label="Answer")
 heading_links  = gr.Markdown(label="Links Heading")
@@ -129,7 +112,7 @@ outputs_chunks = gr.Markdown(label="Used Chunks")
 demo = gr.Interface(
     fn=qa_fn,
     inputs=[
-        gr.Textbox(lines=4, label="What Documents are you looking for?", placeholder="Please change to get propper results:\nDocuments covering the EDPB's stance on automated decision-making, particularly profiling, under the GDPR. Guidelines on how organizations should inform data subjects about automated decisions and the rights of individuals to object to such decisions."),
+        gr.Textbox(lines=4, label="What Documents are you looking for?", placeholder="Please change to get propper results:\nDocuments covering the EDPB’s stance on automated decision-making, particularly profiling, under the GDPR. Guidelines on how organizations should inform data subjects about automated decisions and the rights of individuals to object to such decisions."),
         gr.Textbox(lines=1, label="What is your question?", placeholder="Please change to get propper results:\nWhat does the EDPB recommend regarding automated decision-making and profiling under the GDPR, and what rights do individuals have in relation to such decisions?"),
     ],
     additional_inputs=[