Update app.py

app.py

@@ -1,14 +1,58 @@
 import faiss
 import numpy as np
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import gradio as gr
 
-# Load FAISS index
+# Set paths
 FAISS_PATH = "asa_faiss.index"
-index = faiss.read_index(FAISS_PATH)
+DATASET_PATH = "responses.txt"  # Ensure this file contains indexed responses
+
+# Load FAISS index
+print(f"Loading FAISS index from {FAISS_PATH}...")
+faiss_index = faiss.read_index(FAISS_PATH)
+print("✅ FAISS index loaded successfully!")
+
+# Load dataset responses
+with open(DATASET_PATH, "r", encoding="utf-8") as f:
+    dataset = f.readlines()
+print("✅ Responses dataset loaded!")
+
+# Load model & tokenizer (Ensure model path is correct)
+MODEL_NAME = "Futuresony/my_model"  # Change this if using a local model
+tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+model = AutoModelForCausalLM.from_pretrained(MODEL_NAME)
+
+# Set FAISS threshold (Adjust based on FAISS distance values)
+THRESHOLD = 100  # Lower threshold to improve response accuracy
+
+
+def embed(text):
+    """Convert text to FAISS-compatible vector (Ensure same embeddings as FAISS training)."""
+    tokens = tokenizer.encode(text, add_special_tokens=True)
+    return np.array(tokens, dtype=np.float32).reshape(1, -1)
+
+
+def chatbot_response(user_query):
+    """Fetches response from FAISS or falls back to the model."""
+    query_vector = embed(user_query)  # Convert input to vector
+    D, I = faiss_index.search(query_vector, k=1)  # Search FAISS
+
+    print(f"Closest FAISS match index: {I[0][0]}, Distance: {D[0][0]}")  # Debugging info
+
+    if D[0][0] < THRESHOLD:  # Check if FAISS result is relevant
+        response = dataset[I[0][0]].strip()  # Fetch matched response
+        print("✅ FAISS response used!")
+    else:
+        # Fallback to model-generated response
+        print("⚠️ FAISS match too weak, using model instead.")
+        inputs = tokenizer(user_query, return_tensors="pt")
+        outputs = model.generate(**inputs, max_new_tokens=150)
+        response = tokenizer.decode(outputs[0], skip_special_tokens=True)
 
-# Example query vector (random, replace with actual embedding from your model)
-query_vector = np.random.rand(1, index.d).astype('float32')
+    return response
 
-# Search FAISS index
-D, I = index.search(query_vector, k=1)  # k=1 means get 1 nearest neighbor
 
-print(f"Closest match index: {I[0][0]}, Distance: {D[0][0]}")
+# Gradio UI
+iface = gr.Interface(fn=chatbot_response, inputs="text", outputs="text", title="ASA Microfinance Chatbot")
+iface.launch()