Update app.py

app.py

@@ -1,15 +1,28 @@
 import gradio as gr
 from huggingface_hub import InferenceClient
-from web import search  # Web search for fetching real-time answers
+from web import search  # Import web search tool
 
 client = InferenceClient("Futuresony/future_ai_12_10_2024.gguf")
 
-# Define topics that your model was NOT fine-tuned on
-OUT_OF_SCOPE_TOPICS = ["current events", "latest news", "sports scores", "politics", "celebrity gossip"]
+def is_uncertain(question, response):
+    """Check if the model's response is unreliable."""
+    # 1. If response length is too short, it's likely a guess.
+    if len(response.split()) < 4:
+        return True
 
-def is_out_of_scope(question):
-    """Check if the question relates to topics the model wasn't trained on."""
-    return any(topic in question.lower() for topic in OUT_OF_SCOPE_TOPICS)
+    # 2. If response repeats the question, it might be unsure.
+    if response.lower() in question.lower():
+        return True
+    
+    # 3. If the response contains generic phrases like "Kulingana na utafiti" (According to research)
+    uncertain_phrases = [
+        "Kulingana na utafiti", "Inaaminika kuwa", "Ninadhani", 
+        "It is believed that", "Some people say", "Inasemekana kuwa"
+    ]
+    if any(phrase.lower() in response.lower() for phrase in uncertain_phrases):
+        return True
+    
+    return False
 
 def google_search(query):
     """Fetch search results from Google."""
@@ -38,28 +51,21 @@ def respond(
 
     response = ""
     
-    response_obj = client.chat_completion(
+    for message in client.chat_completion(
         messages,
         max_tokens=max_tokens,
+        stream=True,
         temperature=temperature,
         top_p=top_p,
-        logprobs=True,  # Get log probabilities to measure uncertainty
-    )
-
-    # Extract response text
-    for message in response_obj:
+    ):
         token = message.choices[0].delta.content
         response += token
-        yield response  # Stream response
-
-    # Analyze uncertainty (using log probabilities)
-    logprobs = response_obj.choices[0].logprobs.token_logprobs
-    avg_confidence = sum(logprobs) / len(logprobs) if logprobs else 0
-
-    # If confidence is low OR the question is about out-of-scope topics, use Google
-    if avg_confidence < -5 or is_out_of_scope(message):
+        yield response  # Stream partial responses
+    
+    # If response is unreliable, fetch from Google
+    if is_uncertain(message, response):
         google_response = google_search(message)
-        yield f"🤖 AI (Low confidence): {response}\n\n🌍 Google: {google_response}"
+        yield f"🤖 AI: {response}\n\n🌍 Google: {google_response}"
 
 demo = gr.ChatInterface(
     respond,
@@ -79,4 +85,3 @@ demo = gr.ChatInterface(
 
 if __name__ == "__main__":
     demo.launch()
-