agent.py

from smolagents import (CodeAgent, 
                        GradioUI, 
                        LiteLLMModel, 
                        OpenAIServerModel, 
                        ChatMessage, 
                        ToolCallingAgent)
from smolagents.default_tools import (DuckDuckGoSearchTool, 
                                      VisitWebpageTool, 
                                      WikipediaSearchTool, 
                                     PythonInterpreterTool
                                        )
import yaml
from tools import check_reasoning, ensure_formatting
from tools import (download_youtube_video, youtube_frames_to_images, 
                    use_vision_model, analyze_frames_with_vision,
                         read_file, download_file_from_url, analyze_csv_file, 
                         analyze_excel_file, youtube_transcribe,
                         transcribe_audio)
import os
from dotenv import load_dotenv
import time

load_dotenv()

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

    
class SlowLiteLLMModel(LiteLLMModel):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

    def __call__(self, messages, **kwargs) -> ChatMessage:
        time.sleep(15)
        # prepend onto whatever messages the Agent built
        return super().__call__(messages, **kwargs)


react_model_name = "gemini/gemini-2.5-flash-lite"
react_model = LiteLLMModel(model_id=react_model_name, 
                           api_key=os.getenv("GEMINI_KEY"),
                           temperature=0.2
                           )


search_tool = DuckDuckGoSearchTool()
search_tool.name = "search"

visit_tool = VisitWebpageTool(max_output_length=500000)
visit_tool.name = "open_page"

wiki_tool = WikipediaSearchTool(extract_format='HTML')
wiki_tool.name = "wiki_search"

research_agent = ToolCallingAgent(
    name="ResearchAgent",
    description="""
    Handles web search and information retrieval.

    Tools:
    - search(query: str): Search the web
    - open_page(url: str): Open a webpage
    - wiki_search(query: str): Search Wikipedia
    """,
    model=react_model,
    tools=[search_tool, visit_tool, wiki_tool]
)

vision_tool = use_vision_model
vision_tool.name = "analyze_image"

vision_agent = ToolCallingAgent(
    name="VisionAgent",
    description="""
    Handles image understanding and OCR.

    Tools:
    - extract_text(image_path: str): Extract text from image
    - analyze_image(image_path: str): Analyze image content
    """,
    model=react_model,
    tools=[vision_tool]
)

csv_tool = analyze_csv_file
csv_tool.name = "analyze_csv"

excel_tool = analyze_excel_file
excel_tool.name = "analyze_excel"

data_agent = ToolCallingAgent(
    name="DataAgent",
    description="""
    Handles CSV and Excel analysis.

    Tools:
    - analyze_csv(file_path: str): Analyze CSV file
    - analyze_excel(file_path: str): Analyze Excel file
    """,
    model=react_model,
    tools=[csv_tool, excel_tool]
)

yt_transcribe_tool = youtube_transcribe
yt_transcribe_tool.name = "youtube_transcribe"

yt_frames_tool = youtube_frames_to_images
yt_frames_tool.name = "youtube_frames"

audio_tool = transcribe_audio
audio_tool.name = "transcribe_audio"

analyze_frame = analyze_frames_with_vision
analyze_frame.name = "analyze_frames"

media_agent = ToolCallingAgent(
    name="MediaAgent",
    description="""
    Handles YouTube and audio processing.

    Tools:
    - youtube_transcribe(url: str): Get transcript from YouTube
    - youtube_frames(url: str): Extract frames from downloaded video
    - transcribe_audio(file_path: str): Convert audio to text
    - analyze_frames(url: str): Analyze frames from video
    """,
    model=react_model,
    tools=[
        yt_transcribe_tool,
        yt_frames_tool,
        audio_tool,
        analyze_frame,
    ]
)


manager_agent = CodeAgent(
    model=react_model,
    tools=[ 
           read_file, download_file_from_url, download_youtube_video
           ],
    managed_agents=[
        research_agent,
        vision_agent,
        data_agent,
        media_agent
    ],
    additional_authorized_imports=['os', 'pandas', 'numpy', 'PIL', 'tempfile', 'PIL.Image'],
    max_steps=20,
    verbosity_level=1,
    planning_interval=6,
    name="Manager",
    description="The manager of the team, responsible for overseeing and guiding the team's work. Routes tasks to the right specialist agents",
    #final_answer_checks=[check_reasoning, ensure_formatting],
    prompt_templates=prompt_templates
)



if __name__ == "__main__":
    GradioUI(manager_agent).launch()