agent.py

# agent.py

from llama_index.llms.huggingface_api import HuggingFaceInferenceAPI
from llama_index.core.agent.workflow import AgentWorkflow
from llama_index.core.tools import FunctionTool
from langchain_community.tools import DuckDuckGoSearchRun, WikipediaQueryRun
from langchain_experimental.tools.python.tool import PythonREPLTool
from langchain_community.document_loaders import YoutubeLoader

# Define all tool functions with type annotations

def search_duckduckgo(query: str) -> str:
    """Use DuckDuckGo to search the internet."""
    return DuckDuckGoSearchRun().run(query)

def search_wikipedia(query: str) -> str:
    """Use Wikipedia to look up facts."""
    return WikipediaQueryRun(api_wrapper=None).run(query)

def run_python(code: str) -> str:
    """Execute Python code and return output."""
    return PythonREPLTool().run(code)

def get_youtube_transcript(url: str) -> str:
    """Extract transcript from YouTube video."""
    loader = YoutubeLoader.from_youtube_url(url, add_video_info=False)
    docs = loader.load()
    return " ".join(doc.page_content for doc in docs)

# Build tool wrappers
TOOLS = [
    FunctionTool.from_defaults(search_duckduckgo),
    FunctionTool.from_defaults(search_wikipedia),
    FunctionTool.from_defaults(run_python),
    FunctionTool.from_defaults(get_youtube_transcript),
]

# Load LLM
llm = HuggingFaceInferenceAPI(model_name="Qwen/Qwen2.5-Coder-32B-Instruct")

# Create LlamaIndex agent
agent = AgentWorkflow.from_tools_or_functions(
    TOOLS,
    llm=llm,
    system_prompt="You are a helpful and smart AI agent that solves tasks using reasoning and external tools."
)

# Optional: support context (stateful runs)
from llama_index.core.workflow import Context
ctx = Context(agent)

async def answer_question(question: str) -> str:
    """Run the agent on a single question."""
    return await agent.arun(question)