Update app.py

app.py

@@ -27,6 +27,50 @@ def duckduckgo_search(query):
         results = ddgs.text(query, max_results=10)
     return results
 
+async def rephrase_query(query, history, model):
+    system_message = """You are an AI assistant tasked with analyzing and rephrasing user queries. Your goal is to determine if a query is unique or related to the previous conversation, and then rephrase it appropriately for web search. Keep the rephrased query succinct and in a web search query format.
+
+    If the query is unique, rephrase it to be more specific and searchable.
+    If the query is related to the previous conversation, incorporate relevant context from the previous response.
+
+    Provide your analysis in the following format:
+    <analysis>Your reasoning about whether the query is unique or related</analysis>
+    <rephrased_query>The rephrased query</rephrased_query>"""
+
+    user_message = f"""Current query: {query}
+
+    Previous conversation history:
+    {history}
+
+    Analyze the query and provide a rephrased version suitable for web search."""
+
+    client = InferenceClient(model, token=huggingface_token)
+    
+    try:
+        response = await asyncio.to_thread(
+            client.text_generation,
+            prompt=f"{system_message}\n\n{user_message}",
+            max_new_tokens=150,
+            temperature=0.2,
+        )
+        
+        # Extract the rephrased query from the response
+        analysis_start = response.find("<analysis>")
+        analysis_end = response.find("</analysis>")
+        rephrased_start = response.find("<rephrased_query>")
+        rephrased_end = response.find("</rephrased_query>")
+        
+        if analysis_start != -1 and analysis_end != -1 and rephrased_start != -1 and rephrased_end != -1:
+            analysis = response[analysis_start + 10:analysis_end].strip()
+            rephrased_query = response[rephrased_start + 17:rephrased_end].strip()
+            return analysis, rephrased_query
+        else:
+            logging.error("Failed to parse the rephrase response")
+            return None, query
+    except Exception as e:
+        logging.error(f"Error in rephrase_query: {str(e)}")
+        return None, query
+
 def create_web_search_vectors(search_results):
     embed = get_embeddings()
     documents = []
@@ -36,7 +80,7 @@ def create_web_search_vectors(search_results):
             documents.append(Document(page_content=content, metadata={"source": result['href']}))
     return FAISS.from_documents(documents, embed)
 
-async def get_response_with_search(query, model, use_embeddings, conversation_history, num_calls=3, temperature=0.2):
+async def get_response_with_search(query, model, use_embeddings, num_calls=3, temperature=0.2):
     search_results = duckduckgo_search(query)
     
     if not search_results:
@@ -51,20 +95,14 @@ async def get_response_with_search(query, model, use_embeddings, conversation_hi
     else:
         context = "\n".join([f"{result['title']}\n{result['body']}\nSource: {result['href']}" for result in search_results])
 
-    system_message = """You are a world-class AI system, capable of complex reasoning and reflection. 
+    system_message = """ You are a world-class AI system, capable of complex reasoning and reflection. 
     Reason through the query inside <thinking> tags, and then provide your final response inside <output> tags.
     Providing comprehensive and accurate information based on web search results is essential. 
     Your goal is to synthesize the given context into a coherent and detailed response that directly addresses the user's query. 
     Please ensure that your response is well-structured, factual, and cites sources where appropriate.
-    If you detect that you made a mistake in your reasoning at any point, correct yourself inside <reflection> tags.
-    Consider the conversation history when formulating your response to maintain context and coherence."""
+    If you detect that you made a mistake in your reasoning at any point, correct yourself inside <reflection> tags."""
 
-    conversation_context = "\n".join([f"Human: {msg['human']}\nAI: {msg['ai']}" for msg in conversation_history])
-    
-    user_message = f"""Conversation history:
-{conversation_context}
-
-Using the following context from web search results:
+    user_message = f"""Using the following context from web search results:
 {context}
 
 Write a detailed and complete research document that fulfills the following user request: '{query}'
@@ -110,11 +148,14 @@ async def respond(message, history, model, temperature, num_calls, use_embedding
     logging.info(f"Number of API Calls: {num_calls}")
     logging.info(f"Use Embeddings: {use_embeddings}")
 
-    # Convert Gradio history to a list of dictionaries
-    conversation_history = [{"human": h, "ai": a} for h, a in history]
-
     try:
-        async for main_content, sources in get_response_with_search(message, model, use_embeddings, conversation_history, num_calls=num_calls, temperature=temperature):
+        # Rephrase the query
+        analysis, rephrased_query = await rephrase_query(message, history, model)
+        
+        if analysis:
+            yield f"Query Analysis: {analysis}\n\nRephrased Query: {rephrased_query}\n\nSearching the web...\n\n"
+        
+        async for main_content, sources in get_response_with_search(rephrased_query, model, use_embeddings, num_calls=num_calls, temperature=temperature):
             response = f"{main_content}\n\n{sources}"
             yield response
     except asyncio.CancelledError:
@@ -147,8 +188,8 @@ def create_gradio_interface():
             gr.Slider(minimum=1, maximum=5, value=1, step=1, label="Number of API Calls"),
             gr.Checkbox(label="Use Embeddings", value=False),
         ],
-        title="AI-powered Conversational Web Search Assistant",
-        description="Use web search to answer questions or generate summaries. The assistant remembers previous interactions.",
+        title="AI-powered Web Search Assistant",
+        description="Use web search to answer questions or generate summaries.",
         theme=gr.Theme.from_hub("allenai/gradio-theme"),
         css=css,
         examples=[
@@ -177,7 +218,6 @@ def create_gradio_interface():
         5. Check or uncheck the "Use Embeddings" box to toggle between using embeddings or direct text summarization.
         6. Press Enter or click the submit button to get your answer.
         7. Use the provided examples or ask your own questions.
-        8. The assistant will remember previous interactions and maintain context throughout the conversation.
         """)
 
     return demo