src/app/main_agent.py

from langchain_core.messages import BaseMessage, AIMessage
from langchain_core.runnables import RunnableLambda, Runnable
from langchain_community.llms import Ollama
from langchain.tools import Tool
from langgraph.graph import MessageGraph
import re

llm = Ollama(model="gemma3:1b", temperature=0.0) # llama3.1

def create_agent(accent_tool_obj) -> tuple[Runnable, Runnable]:
    accent_tool = Tool(
        name="AccentAnalyzer",
        func=accent_tool_obj.analyze,
        description="Analyze a public MP4 video URL and determine the English accent with transcription."
    )

    def analyze_node(messages: list[BaseMessage]) -> AIMessage:
        last_input = messages[-1].content
        match = re.search(r'https?://\S+', last_input)
        if match:
            url = match.group()
            result = accent_tool.func(url)
        else:
            result = "No valid video URL found in your message."
        return AIMessage(content=result)

    graph = MessageGraph()
    graph.add_node("analyze_accent", RunnableLambda(analyze_node))
    graph.set_entry_point("analyze_accent")
    graph.set_finish_point("analyze_accent")
    analysis_agent = graph.compile()

    # Follow-up agent that uses transcript and responds to questions
    def follow_up_node(messages: list[BaseMessage]) -> AIMessage:
        user_question = messages[-1].content
        transcript = accent_tool_obj.last_transcript or ""
        prompt = f"""You are given this transcript of a video:

         \"\"\"{transcript}\"\"\"

         Now respond to the user's follow-up question: {user_question}
         """
        response = llm.invoke(prompt)
        return AIMessage(content=response)

    follow_up_agent = RunnableLambda(follow_up_node)

    return analysis_agent, follow_up_agent