app.py

# app.py 
import os
import streamlit as st
from dotenv import load_dotenv 
from langchain_groq import ChatGroq 
from langgraph.graph import StateGraph, START, END
# from langgraph.prebuilt import ToolNode, tools_condition
from langgraph.checkpoint.memory import MemorySaver
from langchain_core.messages import AIMessage, HumanMessage
from langgraph.graph.message import add_messages
from typing import Annotated
from typing_extensions import TypedDict
from langchain_together import Together
from tools import execute_python_code, web_search, deep_think
import io
import contextlib
import traceback
from langchain_core.messages import AIMessage
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain.prompts import PromptTemplate
from langchain.chains import LLMChain
from langchain_core.messages import AIMessage
from typing import List
from langgraph.graph import StateGraph, END
# Load environment
load_dotenv()
# os.environ["TAVILY_API_KEY"] = os.getenv("TAVILY_API_KEY")
os.environ["GROQ_API_KEY"] = os.getenv("GROQ_API_KEY")
together_api_key = os.getenv("TOGETHER_API_KEY")
# Define tools

 
# LangGraph State
class State(TypedDict):
    messages: Annotated[list, add_messages]    
    input : str
    questions : List[str]
    answers:List[str]
    code : str
    explanation:str
    subtasks: List[str]
    follow_up_questions: List[str]
# LLM
code_generator = Together(
    model="deepseek-ai/DeepSeek-R1-Distill-Llama-70B-free",  
    temperature=0.2,
    max_tokens=1500,
    api_key=together_api_key,  # Note: parameter name changed from together_api_key to api_key
)
 
# Memory
memory = MemorySaver()

def generate_questions(state: State):
    user_input = state["messages"][-1].content
    result = subtask_chain.invoke({"user_goal": user_input})
    text = result["text"]
    subtasks, questions = parse_subtasks_and_questions(text)

    follow_up = "\n".join(f"Q{i+1}: {q}" for i, q in enumerate(questions))
    state["messages"].append(AIMessage(content="To proceed, please answer these questions:\n" + follow_up))

    return {
        "messages": state["messages"],
        "questions": questions,
        "answers": [],  # Wait for user
        }

def wait_for_answers(state: State):
    # Just pass through until answers are submitted
    return state
    
def handle_answers(state: State):
    full_input = state["input"] + "\n\n" + "\n".join(state["answers"])
    return {**state, "input": full_input}

# Define node
def ai_assistance(state: State):
    result =code_generator.invoke(state["messages"])
    return {"messages": state['messages']+[result]}

# def agent_node(state: State):
#     # Use your LLM here (e.g., Together, OpenAI, etc.)
#     model = ChatGoogleGenerativeAI(model = "gemini-2.0-flash-001").bind_tools(tools)
#     follow_up_prompt = "Break down this task into subtasks and ask follow-up questions if needed:\n\n"
#     last_user_msg = state["messages"][-1].content
#     full_prompt = follow_up_prompt + last_user_msg

#     response = model.invoke(full_prompt)

#     return {"messages": state["messages"] + [AIMessage(content=response)]}

llm = ChatGroq( model="qwen/qwen3-32b",temperature=0.6)

# Template to extract subtasks from the user's input
subtask_prompt = PromptTemplate.from_template(
    """You are an expert AI agent designer.

Given the user's goal:
"{user_goal}"

1. Break this goal into a clear list of subtasks (in bullet points).
2. If any clarification is needed, ask relevant follow-up questions.

Respond in this format:
---
Subtasks:
- ...
- ...
Follow-Up Questions (if any):
- ...
---"""
)

subtask_chain = LLMChain(llm=llm, prompt=subtask_prompt)

def agent_node(state: State):
    user_input = state["messages"][-1].content

    # Get subtasks and possible questions
    result = subtask_chain.invoke({"user_goal": user_input})
    response_text = result["text"]

    # Parse subtasks and follow-up questions
    subtasks, questions = parse_subtasks_and_questions(response_text)

    # Append AI response to messages
    state["messages"].append(AIMessage(content=response_text))

    # Save subtasks and questions into state
    return {
        "messages": state["messages"],
        "subtasks": subtasks,
        "follow_up_questions": questions
    }

# ✂️ Helper function to parse bullet points
def parse_subtasks_and_questions(text: str):
    subtasks = []
    questions = []

    collecting = None
    for line in text.strip().splitlines():
        line = line.strip()
        if line.lower().startswith("subtasks:"):
            collecting = "subtasks"
        elif line.lower().startswith("follow-up questions"):
            collecting = "questions"
        elif line.startswith("-"):
            if collecting == "subtasks":
                subtasks.append(line[1:].strip())
            elif collecting == "questions":
                questions.append(line[1:].strip())

    return subtasks, questions

import time
def generate_code(state: State):
    user_prompt = state["input"]

    system_prompt = """You are an expert Python coding assistant specializing in LangGraph applications.

    Generate clean, working Python code for the user's request with these requirements:
    1. The code MUST use the LangGraph framework (langgraph library).
    2. Implement a proper flow graph using StateGraph.
    3. Include all necessary imports and make sure the code is complete.
    4. Include code to visualize the flow graph (using builder.show() or similar methods).
    5. Structure the code with proper node functions, state definitions, and graph compilation.
            Your code must include the following:
        1. **LangGraph architecture**: Use StateGraph, add_node, add_edge, set_entry_point, etc.
        2. **Subtask breakdown**: Translate user requirements into multiple graph nodes that represent subtasks.
        3. **LLM Agent**: At least one node should be powered by an LLM (e.g., via langchain or similar).
        4. **Terminal Output**: Include a node that prints or returns the final output.
        5. **Execution Ready**: All necessary imports, type definitions (e.g., TypedDict for state), and execution commands (`graph = builder.compile()` + `graph.invoke()`).

    STRICT RULES:
    - DO NOT explain anything.
    - DO NOT wrap code in markdown.
    - DO NOT add comments.
    IMPORTANT: Output ONLY the final Python code.  
    DO NOT include any explanations, comments, or text before, inside, or after the code.  
    Start the output with the necessary import statements (e.g., "from langgraph import StateGraph, State, Transition").  
    No additional text, no markdown fences, just the pure code.

    User request:"""    
    instruction = f"""
        You are an expert LangGraph developer.
        
        Your task is to generate working Python code using the LangGraph library based on the user's request.
        
        Guidelines:
        - Identify the high-level steps from the user's prompt.
        - Break the task into individual LangGraph nodes (functions).
        - Define a TypedDict for the shared state.
        - Build a `StateGraph` using `add_node`, `add_edge`, and `set_entry_point`.
        - Ensure the graph compiles and ends at the `END` node.
        - Avoid external libraries unless clearly specified.
        - Print final output using a terminal node if needed.
        - Keep it clean, minimal, and executable.
        
        Now, generate the code for this task:
        {user_prompt}
        """

    full_prompt = system_prompt + instruction    
    for attempt in range(3):
        try:
            code_response = code_generator.invoke(full_prompt)
            return {**state,
                    "code": str(code_response)}
        except Exception as e:
            if "503" in str(e):
                print(f"[Retry {attempt+1}/3] Together API unavailable (503). Retrying...")
                time.sleep(2)
            else:
                raise e
    raise Exception("Together API failed after 3 retries.")

def explain_code(state):
    code = state["code"]
    user_prompt = state["input"]

    
    system_prompt = """You are a LangGraph expert who explains code clearly. Provide a detailed explanation of the code in three parts:
    
    1. LANGGRAPH FLOW: Explain the flow graph architecture, including nodes, edges, and how data flows through the graph. Describe what would appear in the flow visualization.
    
    2. CODE FLOW: Explain the high-level flow of the code, its architecture, and how different components interact.
    
    3. CODE EXPLANATION: Break down the code step-by-step so a beginner can understand what each part does.
    
    4. VISUALIZATION INSTRUCTIONS: Provide clear instructions on how to run the code to see the flow visualization.
    
    Make your explanation clear, concise, and educational. Include ASCII art to represent the flow graph if possible.
    
    """
   
    
    prompt = f"""User requested: {user_prompt}
    
    Here's the generated LangGraph code:
    ```python
    {code}
    ```
    
    Explain the LangGraph flow, code architecture, and provide detailed instructions for visualization."""
    
    full_prompt = system_prompt + prompt
    explanation = code_generator.invoke(full_prompt)
    
    return {**state, "explanation": explanation, "code": state.get("code")}

# from langchain.chat_models import ChatOpenAI

# llm = ChatOpenAI(model_name="gpt-4", temperature=0)

# def agent_node(state):
#     input_text = state["input"]
#     result = llm.predict(input_text)
#     return {"response": result}
    
def execute_code(state: State) -> State:
    code = state.get("code", "")
    buffer = io.StringIO()

    try:
        with contextlib.redirect_stdout(buffer):
            exec(code, {})
        output = buffer.getvalue() or "✅ Code executed successfully with no output."
    except Exception:
        output = "❌ Execution Error:\n" + traceback.format_exc()

    return {
        **state,
        "execution_result": output
    }

def subtask_splitter(state):
    input_text = state["input"]
    # Hardcoded LLM call example
    response = llm.predict(f"Split this task into clear LangGraph subtasks:\n{input_text}")
    return {"subtasks": response}


def get_all_tools():
    return [
        # ... other tools
        execute_python_code
    ]

def router(state):
    user_input = state["input"].lower()
    if "generate" in user_input:
        return "Generate_Code"
    else:
        return "AI_Assistance"

# Define your graph builder with the state schema
builder = StateGraph(State)

# Add Nodes
builder.add_node("LLM_Agent", agent_node)
# builder.add_node("AI_Assistance", ai_assistance)
builder.add_node("Generate_Questions", generate_questions)
builder.add_node("Wait_For_Answers", wait_for_answers) 
builder.add_node("Handle_Answers", handle_answers)
 # this must be defined
builder.add_node("Generate_Code", generate_code)
builder.add_node("Code_Explainer", explain_code)

# Set Entry Point
builder.set_entry_point("LLM_Agent")

# Define Conditional Function
def check_if_answered(state: State) -> str:   
    if "answers" in state and state['answers'] and any(state['answers']):
        return "answered"
    else:
        return "not_answered"

# Define Flow
builder.add_edge("LLM_Agent", "Generate_Questions")

builder.add_conditional_edges(
    "Generate_Questions",
    check_if_answered,
    {
        "answered": "Handle_Answers",
        "not_answered": "Wait_For_Answers"
    }
)

builder.add_edge("Wait_For_Answers", "Generate_Questions")
builder.add_edge("Handle_Answers", "Generate_Code")
builder.add_edge("Generate_Code", "Code_Explainer")
builder.add_edge("Code_Explainer", END)

# Optionally: define what happens after waiting (if it's a loop)
# builder.add_edge("Wait_For_Answers", "Generate_Questions")  # retry loop

graph = builder.compile(checkpointer=memory) 
# Streamlit UI setup
st.set_page_config(page_title="MitraVerse", layout="wide")
 
st.markdown("""
<style>
.stChatMessage {
    padding: 12px;
    margin-bottom: 12px;
    border-radius: 12px;
    max-width: 90%;
}
.user {
    background-color: #dcf8c6;
    align-self: flex-end;
}
.bot {
    background-color: #f1f0f0;
    align-self: flex-start;
}
.input-box {
    display: flex;
    align-items: center;
    gap: 0.5rem;
}
#floating-container {
    display: flex;
    align-items: center;
    justify-content: space-between;
    padding: 0.25rem 0.75rem;
    background-color: #f9f9f9;
    border-radius: 0.75rem;
    margin-top: 1rem;
    border: 1px solid #ccc;
}
.floating-popup {
    margin-top: 0.5rem;
    padding: 0.5rem;
    border-radius: 0.5rem;
    border: 1px solid #ccc;
    background-color: white;
}
</style>
""", unsafe_allow_html=True)
 
st.title("🧠MitraVerse")
# Columns for button layout
col1, col2, col3 = st.columns(3) 
# Initialize session
if "thread_id" not in st.session_state:
    st.session_state.thread_id = "1"
if "chat_history" not in st.session_state:
    st.session_state.chat_history = [] 
# Show chat
for msg in st.session_state.chat_history:
    role = "user" if isinstance(msg, HumanMessage) else "bot"
    st.markdown(f"<div class='stChatMessage {role}'>{msg.content}</div>", unsafe_allow_html=True) 

with st.container():
    with st.form("chat_form", clear_on_submit=True):
        st.markdown('<div id="floating-container">', unsafe_allow_html=True)        
        st.markdown('</div>', unsafe_allow_html=True) 

        user_input = st.text_input("Ask me", label_visibility="collapsed", placeholder="Ask me Anything")
        submitted = st.form_submit_button(label="Send") 
        if submitted and user_input:
            st.session_state.chat_history.append(HumanMessage(content=user_input))
            
            config = {"configurable": {"thread_id": st.session_state.thread_id},"recursion_limit" : 50}
            state_input = {
                "messages": st.session_state.chat_history,
                "input": user_input,
                "answers": [],
            }
        
            # First round: check if we already have questions pending
            result = graph.invoke(state_input, config=config)
        
            if result.get("questions") and not result.get("answers"):
                st.session_state.pending_questions = result["questions"]
                st.session_state.latest_state = result  # Save intermediate state
                st.rerun()
            else:
                st.session_state.chat_history = result.get("messages", st.session_state.chat_history)
                if result.get("code"):
                    st.session_state.latest_code = result["code"]
                    st.session_state.chat_history.append(
                        AIMessage(content="**💻 Generated Code:**\n\n```python\n" + result["code"] + "\n```")
                    )
                if result.get("explanation"):
                    st.session_state.latest_explanation = result["explanation"]
                    st.session_state.chat_history.append(
                        AIMessage(content="**🔍 Code Explanation:**\n\n```\n" + result["explanation"] + "\n```")
                    )
                st.rerun()

        elif "pending_questions" in st.session_state and st.session_state.pending_questions:
            st.markdown("### 📝 Please answer the following questions:")
            answers = []
            with st.form("answer_form", clear_on_submit=True):
                for i, question in enumerate(st.session_state.pending_questions):
                    answers.append(st.text_input(f"{question}", key=f"answer_{i}"))
                submit_answers = st.form_submit_button("Submit Answers")
            
            if submit_answers:
                latest_state = st.session_state.latest_state
                latest_state["answers"] = answers
        
                config = {"configurable": {"thread_id": st.session_state.thread_id}}
                result = graph.invoke(latest_state, config=config)
        
                st.session_state.pending_questions = []  # Clear
                st.session_state.chat_history = result.get("messages", st.session_state.chat_history)
        
                if result.get("code"):
                    st.session_state.latest_code = result["code"]
                    st.session_state.chat_history.append(
                        AIMessage(content="**💻 Generated Code:**\n\n```python\n" + result["code"] + "\n```")
                    )
                if result.get("explanation"):
                    st.session_state.latest_explanation = result["explanation"]
                    st.session_state.chat_history.append(
                        AIMessage(content="**🔍 Code Explanation:**\n\n```\n" + result["explanation"] + "\n```")
                    )
        
                st.rerun()
             
                st.markdown("<script>window.scrollTo(0, document.body.scrollHeight);</script>", unsafe_allow_html=True)

# ================================
# TOOL BUTTONS SECTION
# ================================
# col1, col2, col3 = st.columns(3)

# user_prompt = st.session_state.get("latest_code", "") or user_input  # fallback to user_input if needed

# with st.container():
    # if col1.button("⚙️ Run Python Code"):
    #     if user_prompt:
    #         with st.spinner("Executing your Python code..."):
    #             result = execute_python_code.invoke({"code": user_prompt})
    #             st.success("✅ Output:")
    #             st.code(result, language="python")
        # else:
    #     #     st.warning("Please enter Python code in the input box.")

    # if col1.button("🌐 Web Search"):
    #     if user_prompt:
    #         with st.spinner("Searching the web..."):
    #             result = web_search.invoke({"query": user_prompt})
    #             st.success("🔎 Search Result:")
    #             st.write(result)
    #     else:
    #         st.warning("Please enter a search query.")

    # if col2.button("🧠 Deep Think"):
    #     if user_prompt:
    #         with st.spinner("Thinking deeply..."):
    #             result = deep_think.invoke({"prompt": user_prompt})
    #             st.success("🧠 Reasoned Output:")
    #             st.write(result)
    #     else:
    #         st.warning("Please enter a prompt.")