app.py

import yaml
import re
from smolagents import CodeAgent, HfApiModel
from tools.final_answer import FinalAnswerTool
from Gradio_UI import GradioUI

# Updated system prompt: Only output the final, direct advice in plain text.
system_prompt = (
    "You are a health and lifestyle advisor specializing in the early detection and prevention of hypertension. "
    "Provide only the final, direct, and concise lifestyle advice based solely on the user's details. "
    "Do NOT include any internal reasoning, chain-of-thought, intermediate steps, or code snippets. "
    "Output exactly one final answer as plain text with no extra commentary."
)

def remove_code_snippets(text):
    """
    Removes code blocks, inline code, chain-of-thought, and debugging/step logs from the output.
    """
    # Remove triple-backtick code blocks.
    text = re.sub(r"```[\s\S]+?```", "", text, flags=re.DOTALL)
    # Remove inline code enclosed in single backticks.
    text = re.sub(r"`[^`]+`", "", text)
    # Remove any text between <think> and </think> tags.
    text = re.sub(r"<think>[\s\S]*?</think>", "", text, flags=re.DOTALL)
    # Remove debug/step log banners (e.g., "━━━━━ Step X ━━━━━")
    text = re.sub(r"━+.*Step \d+.*━+", "", text)
    # Remove any lines that start with "[Step" (which include duration and token info).
    text = re.sub(r"\[Step \d+: Duration .*", "", text)
    # Remove lines that mention code snippet instructions.
    text = re.sub(r"Make sure to include code with the correct pattern.*", "", text)
    # Finally, remove any remaining lines that seem to be debug logs.
    lines = text.splitlines()
    cleaned_lines = [line for line in lines if not re.search(r"Step \d+|Duration", line)]
    return "\n".join(cleaned_lines).strip()

# Use only the final_answer tool.
final_answer = FinalAnswerTool()

# Set up the model with a reduced token limit.
model = HfApiModel(
    max_tokens=1024,
    temperature=0.5,
    model_id='deepseek-ai/DeepSeek-R1-Distill-Qwen-32B',
    custom_role_conversions=None,
)

# Load prompt templates from YAML.
with open("prompts.yaml", 'r') as stream:
    prompt_templates = yaml.safe_load(stream)

# Ensure the final_answer key exists in prompt_templates to prevent KeyError.
if "final_answer" not in prompt_templates:
    prompt_templates["final_answer"] = {"pre_messages": "", "post_messages": ""}

# Initialize CodeAgent with a low verbosity level to reduce extra debug output.
agent = CodeAgent(
    model=model,
    tools=[final_answer],
    verbosity_level=0,
    grammar=None,
    planning_interval=None,
    name="Hypertension Prevention Advisor",
    description=system_prompt,
    prompt_templates=prompt_templates
)

def run_agent(user_input):
    """
    Runs the agent, then removes any internal chain-of-thought, step logs, and code snippets
    before returning the final plain-text answer.
    """
    raw_response = agent.run(user_input)
    print("Raw Agent Response:", raw_response)  # Debugging output (optional)
    
    if not raw_response.strip():
        return "I'm sorry, but I couldn't generate a response. Please try again."
    if "final_answer" not in raw_response.lower():
        return "Error: The response did not use the `final_answer` tool. Please try again."
    
    clean_response = remove_code_snippets(raw_response)
    words = clean_response.split()
    if len(set(words)) < 5:
        return "I'm unable to generate a meaningful response. Please refine your query."
    return clean_response

# Launch the Gradio UI.
GradioUI(agent).launch()