backend/llm_agent.py

from langgraph.graph import StateGraph, END
from langchain_core.messages import HumanMessage, SystemMessage
from langchain_openai import ChatOpenAI  # or your preferred LLM
from pydantic import BaseModel, Field, field_validator
from typing import TypedDict, List
import os
from pydantic import field_validator  # needed for custom validation
from dotenv import load_dotenv
load_dotenv()  # take environment variables from .env.

# Define the structured output model
class StoryOutput(BaseModel):
    """Structured output for the storyteller agent"""
    polished_story: str = Field(
        description="A refined version of the story with improved flow, grammar, and engagement"
    )
    keywords: List[str] = Field(
        description="A list of 5-10 key terms that represent the main themes, characters, or concepts",
        min_items=3,
        max_items=7
    )

# Define the state structure
class AgentState(TypedDict):
    original_story: str
    polished_story: str
    keywords: List[str]
    messages: List[dict]

# Storyteller Agent with Structured Output
class StorytellerAgent:
    def __init__(self, llm):
        # Create structured LLM with the output model
        self.structured_llm = llm.with_structured_output(StoryOutput)
        self.system_prompt = """You are a skilled storyteller AI. Your job is to take raw, confessional-style stories and transform them into emotionally engaging, narrative-driven pieces. The rewritten story should:

1. Preserve the original events and meaning but present them in a captivating way.
2. Use character names (instead of “my brother,” “my sister”) to make the story feel alive.
3. Add dialogue, atmosphere, and inner thoughts to create tension and immersion.
4. Write in a third-person narrative style, as if the story is being shared by an observer.
5. Maintain a natural, human voice — conversational, reflective, and vivid.
6. Balance realism with storytelling techniques (scene-setting, emotional beats, sensory details).
7. Keep the length roughly 2–3x the original input, ensuring it feels like a polished story.

Your goal is to make the reader feel emotionally invested, as though they’re listening to someone recounting a deeply personal and dramatic life event.

"""

    def __call__(self, state: AgentState) -> AgentState:
        # Prepare messages for the structured LLM
        messages = [
            SystemMessage(content=self.system_prompt),
            HumanMessage(content=f"Please polish this story and extract keywords:\n\n{state['original_story']}")
        ]
        
        # Get structured response
        response: StoryOutput = self.structured_llm.invoke(messages)
        
        # Update state with structured output
        state["polished_story"] = response.polished_story
        state["keywords"] = response.keywords
        state["messages"].append({
            "role": "assistant", 
            "content": f"Polished story and extracted {len(response.keywords)} keywords"
        })
        
        return state

# Create the graph functions
def create_storyteller_graph(enhanced=False):
    llm = ChatOpenAI(
        model='gpt-4o',
        api_key=os.getenv('OPENAI_API_KEY'),
        temperature=0.2,
        max_tokens=10000
    )

    # Choose agent type
    storyteller = StorytellerAgent(llm)
    
    # Create the graph
    workflow = StateGraph(AgentState)
    workflow.add_node("storyteller", storyteller)
    workflow.set_entry_point("storyteller")
    workflow.add_edge("storyteller", END)
    
    return workflow.compile()

# Usage functions
def process_story(original_story: str, enhanced=False):
    graph = create_storyteller_graph(enhanced)
    
    initial_state = {
        "original_story": original_story,
        "polished_story": "",
        "keywords": [],
        "messages": []
    }
    
    result = graph.invoke(initial_state)
    
    return {
        "polished_story": result["polished_story"],
        "keywords": result["keywords"]
    }

# Example with validation
class ValidatedStoryOutput(BaseModel):
    """Story output with additional validation"""
    polished_story: str = Field(
        description="Enhanced story",
        min_length=50  # Ensure minimum story length
    )
    keywords: List[str] = Field(
        description="Story keywords",
        min_items=3,
        max_items=7
    )

    @field_validator('polished_story')
    def validate_story_quality(cls, v: str):
        """Custom validation for story content"""
        if len(v.split()) < 10:
            raise ValueError("Polished story must contain at least 10 words")
        return v