demo.py

import gradio as gr
import constants as cte
from rag_smolagent import SmolAgent, MessageRole
from googlesearch import search
from typing import List, Tuple, Optional, Any

class ChatState:
    def __init__(self) -> None:
        self.sources: str = "There are no sources yet."
        self.extra_info: str = ""
        self.age: str = ""
        self.residence: str = ""
        self.response_type: str = "Concise"  # Default response type
        self.seen_messages: List[Any] = []
        self.cot_steps: str = ""

state = ChatState()

# Initialize the SmolAgent with the specified data path
agent = SmolAgent()

def get_first_result(query: str) -> Optional[str]:
    """

    Searches Google using a processed version of the query string and returns the URL of the first result.

    """
    parts = query.split(",")
    if len(parts) > 3:
        msg = parts[-3] + parts[-2] + " manual de la renta 2023"   # search only in 2023's Manual
    else:
        msg = parts[-2] + " manual de la renta 2023"

    try:
        # Perform the search (updated configuration)
        results = search(
            msg,
            num_results=1,   # Desired number of results
            lang="es",       # Search language (optional)
        )
        
        return next(results) if results else None
        
    except Exception as e:
        print(f"Error: {e}")
        return None

def process_sources(sources: str) -> str:
    """

    Processes a string with multiple sources and generates Markdown links using the first search result for each source.

    """
    lines: List[str] = sources.strip().split('\n')
    results: List[str] = []
    
    for line in lines:
        clean_line = line.strip()
        if not clean_line:
            continue
        url = get_first_result(clean_line)
        url_str = url + "\n" if url else "Sin resultados"
        results.append(f"[{clean_line}]({url_str})")
    
    return '\n'.join(results)

def update_age(value: str) -> None:
    """Updates the age in the state and refreshes the extra info."""
    state.age = value if value else ""
    update_extra_info()

def update_residence(value: str) -> None:
    """Updates the residence in the state and refreshes the extra info."""
    state.residence = value if value else ""
    update_extra_info()

def update_extra_info() -> None:
    """Updates the extra info by combining age and residence."""
    if state.age and state.residence:
        state.extra_info = f"Tengo {state.age} años y resido en {state.residence}."
    else:
        state.extra_info = ""

def update_response_type(value: str) -> None:
    """Updates the response type in the state."""
    state.response_type = value

def get_prompt(query: str, extra_info: str, response_type: str) -> str:
    """Injects the user's query, extra info, and response type into the prompt template."""
    return cte.PROMPT_TEMPLATE.format(query=query, extra_info=extra_info, response_type=response_type)

def chatbot_response(message: str, history: List[Tuple[str, str]]) -> str:
    """Calls the agent, separates the main response, sources, and extracts CoT steps."""
    response: str = agent(get_prompt(message, state.extra_info, state.response_type))

    # Extract sources if present
    if "Fuentes:" in response:
        state.sources = response.split("Fuentes:")[1].strip()
        response = response.split("Fuentes:")[0].strip()
    else:
        state.sources = "No sources were used for the generation of this message"

    # Extract the main response
    if "Respuesta:" in response:
        answer: str = response.split("Respuesta:")[1].strip()
    else:
        answer = response

    # Extract Chain of Thought (CoT) steps
    cot_messages: List[Any] = agent.agent.write_memory_to_messages()
    state.cot_steps = ""
    cot_step_counter: int = 0

    for message in cot_messages:
        if (message.get('role') == MessageRole.ASSISTANT) & (message not in state.seen_messages):
            for content in message.get('content', []):
                if content.get('type') == 'text' and 'Thought:' in content.get('text', ''):
                    state.cot_steps += f"\n==============================    Step {cot_step_counter}    ==============================\n"
                    cot_step_counter += 1
                    state.cot_steps += (content['text'] + "\n")
                    state.seen_messages.append(message)

    return answer.strip()

def respond(message, chat_history):
    """Handles user input, generates a response, and updates the chat history."""
    chat_history.append((message, ""))
    yield "", chat_history
    
    chat_history[-1] = (message, "⏳ Procesando respuesta...")
    yield "", chat_history
    
    bot_message: str = chatbot_response(message, chat_history)
    
    # Append sources and Chain of Thought as accordions if available
    if state.sources != "No sources were used for the generation of this message":
        try:
            all_url = process_sources(state.sources)
        except Exception:
            all_url = state.sources
        accordion_sources: str = (
            "\n\n<details>\n"
            "  <summary style='color: #0090ff; cursor: pointer;'>Sources</summary>\n\n"
            f"  {all_url}\n\n"
            "</details>"
        )
        bot_message += accordion_sources

    if state.cot_steps:
        accordion_cot: str = (
            "\n\n<details>\n"
            "  <summary style='color: #0090ff; cursor: pointer;'>Chain of Thought</summary>\n\n"
            f"  {state.cot_steps}\n\n"
            "</details>"
        )
        bot_message += accordion_cot

    chat_history[-1] = (message, bot_message)
    yield "", chat_history

def clear_history(chat_history: List[Tuple[str, str]]) -> List[Tuple[str, str]]:
    """Clears the chat history and the agent's internal memory."""
    chat_history.clear()
    global agent
    agent = SmolAgent()
    state.sources = "There are no sources yet."
    state.seen_messages.clear()
    return chat_history

# Gradio UI setup
custom_theme = gr.Theme.load("/teamspace/studios/this_studio/AgenticRAG/theme.json")

with gr.Blocks(
    fill_height=True,
    fill_width=True,
    theme=custom_theme
    ) as demo:
    
    ## Demo Sidebar
    with gr.Sidebar():
        gr.Markdown("## User Information")
        age_input = gr.Textbox(label="Age", placeholder="Unknown", value=state.age)
        residence_input = gr.Textbox(label="Residence", placeholder="Unknown", value=state.residence)
        
        # Update values when fields change
        age_input.change(update_age, inputs=age_input)
        residence_input.change(update_residence, inputs=residence_input)
        
        gr.Markdown("## Answer customization")
        # Dropdown for response type
        response_type_dropdown = gr.Dropdown(
            label="Response Type",
            choices=["Concise", "Detailed"],
            value="Concise"
        )
        response_type_dropdown.change(update_response_type, inputs=response_type_dropdown)
        
        # Clear History button
        clear_button = gr.Button("Clear History", variant="secondary")

        web_link = gr.HTML("<a href='https://sede.agenciatributaria.gob.es/Sede/Ayuda/23Manual/100.html' target='_blank'>Abrir página web</a>")
    
    gr.Markdown("# RAG ChatBot Agent")
    
    with gr.Column(scale=1):
        chatbot = gr.Chatbot(
            label="Chat", 
            bubble_full_width=True, 
            scale=1,
            avatar_images=(None, "https://logosandtypes.com/wp-content/uploads/2022/03/cognizant.svg"),
            group_consecutive_messages=False
        )
                
        msg = gr.Textbox(
            show_label=False,
            placeholder="What do you want to know?",
            submit_btn=True 
        )
    
    # Submit and clear history functionality
    msg.submit(respond, [msg, chatbot], [msg, chatbot])
    clear_button.click(clear_history, inputs=chatbot, outputs=chatbot)

demo.launch(share=True)