Update app.py

app.py

@@ -3,31 +3,58 @@ import pandas as pd
 import numpy as np
 from transformers import pipeline, BertTokenizer, BertModel
 import faiss
+import torch
+import json
 import spaces
 
 # Load CSV data
 data = pd.read_csv('RBDx10kstats.csv')
 
-# Convert embedding column from string to numpy array
-data['embedding'] = data['embedding'].apply(lambda x: np.fromstring(x[1:-1], sep=', '))
+# Function to safely convert JSON strings to numpy arrays
+def safe_json_loads(x):
+    try:
+        return np.array(json.loads(x), dtype=np.float32)  # Ensure the array is of type float32
+    except json.JSONDecodeError as e:
+        print(f"Error decoding JSON: {e}")
+        return np.array([], dtype=np.float32)  # Return an empty array or handle it as appropriate
+
+# Apply the safe_json_loads function to the embedding column
+data['embedding'] = data['embedding'].apply(safe_json_loads)
+
+# Filter out any rows with empty embeddings
+data = data[data['embedding'].apply(lambda x: x.size > 0)]
 
 # Initialize FAISS index
-dimension = len(data['embedding'][0])
-index = faiss.IndexFlatL2(dimension)
-index.add(np.stack(data['embedding'].values))
+dimension = len(data['embedding'].iloc[0])
+res = faiss.StandardGpuResources()  # use a single GPU
 
-# Load QA model
-qa_model = pipeline("question-answering", model="distilbert-base-uncased-distilled-squad")
+# Create FAISS index
+if faiss.get_num_gpus() > 0:
+    gpu_index = faiss.IndexFlatL2(dimension)
+    gpu_index = faiss.index_cpu_to_gpu(res, 0, gpu_index)  # move to GPU
+else:
+    gpu_index = faiss.IndexFlatL2(dimension)  # fall back to CPU
 
-# Function to embed the question using BERT
-def embed_question(question, model, tokenizer):
-    inputs = tokenizer(question, return_tensors='pt')
-    outputs = model(**inputs)
-    return outputs.last_hidden_state.mean(dim=1).detach().numpy()
+# Ensure embeddings are stacked as float32
+embeddings = np.vstack(data['embedding'].values).astype(np.float32)
+gpu_index.add(embeddings)
+
+# Check if GPU is available
+device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+# Load QA model
+qa_model = pipeline("question-answering", model="distilbert-base-uncased-distilled-squad", device=0 if torch.cuda.is_available() else -1)
 
 # Load BERT model and tokenizer
 tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-model = BertModel.from_pretrained('bert-base-uncased')
+model = BertModel.from_pretrained('bert-base-uncased').to(device)
+
+# Function to embed the question using BERT
+def embed_question(question, model, tokenizer):
+    inputs = tokenizer(question, return_tensors='pt').to(device)
+    with torch.no_grad():
+        outputs = model(**inputs)
+    return outputs.last_hidden_state.mean(dim=1).cpu().numpy().astype(np.float32)
 
 # Function to retrieve the relevant document and generate a response
 @spaces.GPU(duration=120)
@@ -36,7 +63,7 @@ def retrieve_and_generate(question):
     question_embedding = embed_question(question, model, tokenizer)
     
     # Search in FAISS index
-    _, indices = index.search(question_embedding, k=1)
+    _, indices = gpu_index.search(question_embedding, k=1)
     
     # Retrieve the most relevant document
     relevant_doc = data.iloc[indices[0][0]]
@@ -50,7 +77,7 @@ def retrieve_and_generate(question):
 # Create a Gradio interface
 interface = gr.Interface(
     fn=retrieve_and_generate,
-    inputs=gr.inputs.Textbox(lines=2, placeholder="Ask a question about the documents..."),
+    inputs=gr.Textbox(lines=2, placeholder="Ask a question about the documents..."),
     outputs="text",
     title="RAG Chatbot",
     description="Ask questions about the documents in the CSV file."