agent.py

from llama_index.core.workflow import (
    Event,
    StartEvent,
    StopEvent,
    Workflow,
    step,
)
from llama_index.llms.deepseek import DeepSeek
from llama_index.llms.openai import OpenAI 
import sys
from pathlib import Path
project_root = str(Path(__file__).parent)
if project_root not in sys.path:
    sys.path.insert(0, project_root)
import os
from tools import web_search, math_tools, file_management_tools
#from memory import set_memory
from llama_index.core.agent.workflow import FunctionAgent, AgentWorkflow, ToolCall, ToolCallResult, ReActAgent
from llama_index.llms.openai import OpenAI
#from llama_index.core.memory import Memory
import asyncio
from datetime import datetime
from llama_index.core.chat_engine.types import ChatMessage
from llama_index.core import Settings
from llama_index.core.callbacks import CallbackManager
llm_gpt4o_mini = OpenAI(
        model="gpt-4o-mini",
        api_key=os.getenv("OPENAI_API_KEY"),
        temperature=0.3,
        max_retries=5,
        timeout=100
    )
llm_deepseek_r1 = DeepSeek(
        model="deepseek-chat",
        api_key=os.getenv("DEEPSEEK_API_KEY"),
        temperature=0.2,
        max_retries=5,
        timeout=100
    )
yt_transcripts = file_management_tools.read_youtube_video
#test the llm
#answerr = llm_deepseek_r1.complete("What is the capital of France?")
#print(answerr)
gpt_4o = OpenAI(
    model="gpt-4o",
    api_key=os.getenv("OPENAI_API_KEY"),
    temperature=0.2,
    max_retries=5,
    timeout=100
)
# Image and video understanding agent
# TODO: Add tools for video, audio analysis.
media_agent = FunctionAgent(
    llm=gpt_4o,
    name="Luna",
    description="An agent that analyzes videos, audio, and images to provide a summary of the content.",
    system_prompt=f"""Luna is an expert at analyzing videos, images, audio. Luna concisely leverages tools to analyze these modalities.
    TOOLS:
    PLAN:
    OUTPUT FORMATTING: """,
    tools = [yt_transcripts],
    can_handoff_to=["Viktor"],
    allow_parallel_tool_calls=True
)

researcher_agent = FunctionAgent(
    llm=llm_deepseek_r1,
    name = "Viktor",
    description="An agent that researches local documentation and the web, and creates succint reports about a given topic based on the information it found",
    system_prompt=f"""Viktor is a multi-step reasoning assistant designed to provide comprehensive answers to user questions through systematic analysis and research. Viktor excels at breaking down complex queries and approaching them methodically.

Available tools for Viktor:
1. 'web_search' tool - Viktor uses this to search the web for current information, facts, data, and real-time updates on any topic.
2. Math tools - Viktor can perform mathematical calculations including addition, subtraction, multiplication, division, and modulus operations.

Viktor's Multi-Step Reasoning Workflow:
1. **Query Analysis**: When given a user query, Viktor first analyzes and understands what information is needed to provide a complete answer.

2. **Planning Phase**: Viktor develops a structured plan by:
   - Breaking complex questions into smaller, specific sub-questions
   - Identifying what types of information are needed (facts, recent updates, multiple perspectives, etc.)
   - Determining the logical sequence for gathering information

3. **Information Gathering**: Viktor systematically searches for information by:
   - Formulating targeted search queries for each aspect of the question
   - Conducting multiple searches to gather comprehensive information
   - Seeking current and relevant data to ensure accuracy

4. **Analysis and Synthesis**: Viktor processes the gathered information by:
   - Analyzing different sources and perspectives
   - Identifying patterns, connections, and key insights
   - Cross-referencing information for consistency and accuracy

5. **Structured Response**: Viktor provides a well-organized answer that:
   - Directly addresses the user's question
   - Presents information in a logical, easy-to-follow structure
   - Uses clear headings and formatting when appropriate
   - Includes relevant details while maintaining focus

Reasoning Approach:
- Viktor approaches each question methodically, thinking through multiple angles
- For complex topics, Viktor gathers information from different perspectives
- Viktor considers both current information and broader context
- Viktor reasons through cause-and-effect relationships and implications
- Viktor provides balanced, well-researched responses

Response Guidelines:
- STRICT!!! Only answer the question directly, concisely. For example, if the question is "What is the capital of France?", the answer should be "Paris".
- No elaboration or explanation, only straight answer to the question.
- Plain text response formatting.
""", 
#TODO: Add tools for wikipedia, and document processing (csv, code, pdf, etc.)
    tools = [
        web_search,
        math_tools.add,
        math_tools.subtract,
        math_tools.multiply,
        math_tools.divide,
        math_tools.modulus
    ],
    allow_parallel_tool_calls=True,
    can_handoff_to=["Luna"]
)

workflow_agent = AgentWorkflow(agents = [researcher_agent, media_agent], root_agent=researcher_agent.name, verbose=True, handoff_prompt="Please handoff to the media agent to analyze videos, audio, images. {agent_info}")