test / chatbot_langgraph_validated.py
add
e983a89 verified 12 months ago
31.3 kB
"""
LangGraph-based Car Finder Chatbot with Validation Agents

This implementation includes:
- Template-based SQL queries (no SQL generation)
- LangGraph for agent orchestration
- Tool-based architecture for security
- Content moderation agent (validates user input)
- Response quality agent (validates assistant responses)
- State management for conversation flow
"""

from typing import TypedDict, Annotated, Optional, Literal
from operator import add
import sqlite3
import os
from dotenv import load_dotenv

from langchain_openai import ChatOpenAI
from langchain_core.messages import HumanMessage, AIMessage, SystemMessage
from langchain_core.tools import tool
from langgraph.graph import StateGraph, END
from langgraph.prebuilt import ToolNode
from pydantic import BaseModel, Field

# Load environment variables
load_dotenv()

# Validate API key
api_key = os.environ.get("OPENAI_API_KEY")
if not api_key:
    raise ValueError("OPENAI_API_KEY environment variable is not set")

# Initialize LLMs
llm = ChatOpenAI(model="gpt-4o", temperature=0.7, api_key=api_key)
llm_validator = ChatOpenAI(model="gpt-4o-mini", temperature=0, api_key=api_key)  # Cheaper for validation

# Load database schema with error handling
try:
    with open('database_schema.txt', 'r') as f:
        SCHEMA_DESCRIPTION = f.read()
except FileNotFoundError:
    raise FileNotFoundError("database_schema.txt not found. Please ensure it exists in the current directory.")

# Constants
MIN_RESULTS = 1
MAX_RESULTS = 20
DB_PATH = 'cars.db'


# ============================================================================
# PYDANTIC MODELS FOR VALIDATION
# ============================================================================

class ContentModerationResult(BaseModel):
    """Result from content moderation validation"""
    is_safe: bool = Field(description="True if content is safe, False if harmful")
    violation_type: Optional[Literal["hate_speech", "harassment", "profanity", "spam", "inappropriate", "off_topic"]] = None
    severity: Optional[Literal["low", "medium", "high"]] = None
    explanation: str = Field(description="Brief explanation of the decision")
    should_block: bool = Field(description="True if the message should be blocked from processing")


class ResponseQualityResult(BaseModel):
    """Result from response quality validation"""
    is_relevant: bool = Field(description="True if response is relevant to user query")
    stays_in_role: bool = Field(description="True if response stays in car shopping assistant role")
    is_helpful: bool = Field(description="True if response is helpful and actionable")
    issues_found: list[str] = Field(default_factory=list, description="List of quality issues")
    severity: Optional[Literal["minor", "major", "critical"]] = None
    explanation: str = Field(description="Brief explanation of quality assessment")
    should_regenerate: bool = Field(description="True if response should be regenerated")


class SearchParameters(BaseModel):
    """Parameters for searching cars using template-based SQL"""
    min_price: Optional[int] = Field(None, description="Minimum price in USD", ge=0, le=100000)
    max_price: Optional[int] = Field(None, description="Maximum price in USD", ge=0, le=100000)
    fuel_type: Optional[Literal["Gasoline", "Diesel", "Electric", "Hybrid", "Plug-in Hybrid"]] = Field(None, description="Type of fuel")
    is_suv: Optional[bool] = Field(None, description="True for SUVs, False for sedans/coupes")
    min_seating: Optional[int] = Field(None, description="Minimum seating capacity", ge=4, le=8)
    max_seating: Optional[int] = Field(None, description="Maximum seating capacity", ge=4, le=8)
    drivetrain: Optional[Literal["FWD", "RWD", "AWD", "4WD"]] = Field(None, description="Drive system")
    min_fuel_efficiency_city: Optional[float] = Field(None, description="Minimum city MPG", ge=0)
    min_cargo_space: Optional[int] = Field(None, description="Minimum cargo space in cubic feet", ge=0)
    has_sunroof: Optional[bool] = Field(None, description="Must have sunroof")
    has_leather_seats: Optional[bool] = Field(None, description="Must have leather seats")
    has_navigation: Optional[bool] = Field(None, description="Must have navigation system")
    has_backup_camera: Optional[bool] = Field(None, description="Must have backup camera")
    min_safety_rating: Optional[float] = Field(None, description="Minimum safety rating", ge=0, le=5)


# ============================================================================
# VALIDATION TOOLS
# ============================================================================

@tool
def moderate_user_input(user_message: str) -> dict:
    """
    Validates user input for harmful speech, spam, or off-topic content.

    Checks for:
    - Hate speech, harassment, discrimination
    - Profanity or vulgar language
    - Spam or scam attempts
    - Off-topic requests (not about cars)
    - Incoherent or nonsensical input

    Returns moderation result with safety verdict.
    """
    system_prompt = """Moderate user input for a car chatbot. Flag only serious issues.

Block if: hate speech, harassment, spam, sexual content, threats, illegal requests, complete gibberish.

Allow: greetings, car questions, normal conversation.

Return is_safe (true/false) and should_block (true only if harmful)."""

    try:
        response = llm_validator.with_structured_output(ContentModerationResult).invoke([
            SystemMessage(content=system_prompt),
            HumanMessage(content=f"Moderate this user message:\n\n{user_message}")
        ])

        return {
            "is_safe": response.is_safe,
            "violation_type": response.violation_type,
            "severity": response.severity,
            "explanation": response.explanation,
            "should_block": response.should_block
        }
    except Exception as e:
        # On error, default to safe (don't block)
        return {
            "is_safe": True,
            "violation_type": None,
            "severity": None,
            "explanation": f"Moderation check failed: {str(e)}",
            "should_block": False
        }


@tool
def validate_assistant_response(user_message: str, assistant_response: str) -> dict:
    """
    Validates that the assistant's response is relevant, helpful, and stays in role.

    Checks:
    - Relevance to user's question
    - Stays in car shopping assistant role
    - Provides actionable information
    - No hallucinations or false claims
    - Professional and helpful tone

    Returns quality assessment with regeneration recommendation.
    """
    system_prompt = """Check if assistant response is relevant, helpful, and stays in car assistant role.

Regenerate only if seriously flawed (off-topic, unhelpful, incoherent).

Minor issues are OK."""

    try:
        response = llm_validator.with_structured_output(ResponseQualityResult).invoke([
            SystemMessage(content=system_prompt),
            HumanMessage(content=f"""Validate this conversation:

USER: {user_message}

ASSISTANT: {assistant_response}

Assess the assistant's response quality.""")
        ])

        return {
            "is_relevant": response.is_relevant,
            "stays_in_role": response.stays_in_role,
            "is_helpful": response.is_helpful,
            "issues_found": response.issues_found,
            "severity": response.severity,
            "explanation": response.explanation,
            "should_regenerate": response.should_regenerate
        }
    except Exception as e:
        # On error, assume response is acceptable
        return {
            "is_relevant": True,
            "stays_in_role": True,
            "is_helpful": True,
            "issues_found": [],
            "severity": None,
            "explanation": f"Validation check failed: {str(e)}",
            "should_regenerate": False
        }


# ============================================================================
# SQL TEMPLATE-BASED TOOLS (Same as before)
# ============================================================================

@tool
def search_cars(
    min_price: Optional[int] = None,
    max_price: Optional[int] = None,
    fuel_type: Optional[str] = None,
    is_suv: Optional[bool] = None,
    min_seating: Optional[int] = None,
    max_seating: Optional[int] = None,
    drivetrain: Optional[str] = None,
    min_fuel_efficiency_city: Optional[float] = None,
    min_cargo_space: Optional[int] = None,
    has_sunroof: Optional[bool] = None,
    has_leather_seats: Optional[bool] = None,
    has_navigation: Optional[bool] = None,
    has_backup_camera: Optional[bool] = None,
    min_safety_rating: Optional[float] = None,
) -> dict:
    """Search for cars using a secure template-based SQL query."""
    conditions = []
    params = []

    if min_price is not None:
        conditions.append("price >= ?")
        params.append(min_price)
    if max_price is not None:
        conditions.append("price <= ?")
        params.append(max_price)
    if fuel_type is not None:
        conditions.append("fuel_type = ?")
        params.append(fuel_type)
    if is_suv is not None:
        conditions.append("is_suv = ?")
        params.append(1 if is_suv else 0)
    if min_seating is not None:
        conditions.append("seating_capacity >= ?")
        params.append(min_seating)
    if max_seating is not None:
        conditions.append("seating_capacity <= ?")
        params.append(max_seating)
    if drivetrain is not None:
        conditions.append("drivetrain = ?")
        params.append(drivetrain)
    if min_fuel_efficiency_city is not None:
        conditions.append("fuel_efficiency_city >= ?")
        params.append(min_fuel_efficiency_city)
    if min_cargo_space is not None:
        conditions.append("cargo_space >= ?")
        params.append(min_cargo_space)
    if has_sunroof is not None:
        conditions.append("has_sunroof = ?")
        params.append(1 if has_sunroof else 0)
    if has_leather_seats is not None:
        conditions.append("has_leather_seats = ?")
        params.append(1 if has_leather_seats else 0)
    if has_navigation is not None:
        conditions.append("has_navigation = ?")
        params.append(1 if has_navigation else 0)
    if has_backup_camera is not None:
        conditions.append("has_backup_camera = ?")
        params.append(1 if has_backup_camera else 0)
    if min_safety_rating is not None:
        conditions.append("safety_rating >= ?")
        params.append(min_safety_rating)

    where_clause = " AND ".join(conditions) if conditions else "1=1"
    query = f"SELECT * FROM cars WHERE {where_clause} ORDER BY price LIMIT 21"

    try:
        with sqlite3.connect(DB_PATH) as conn:
            conn.row_factory = sqlite3.Row
            cursor = conn.cursor()
            cursor.execute(query, params)
            results = cursor.fetchall()

            cars = [dict(row) for row in results]
            count = len(cars)

            if count < MIN_RESULTS:
                status = "too_few"
            elif count > MAX_RESULTS:
                status = "too_many"
                cars = cars[:MAX_RESULTS]
            else:
                status = "good"

            return {
                "count": count,
                "cars": cars,
                "status": status,
                "params_used": {k: v for k, v in [
                    ("min_price", min_price), ("max_price", max_price),
                    ("fuel_type", fuel_type), ("is_suv", is_suv),
                    ("min_seating", min_seating), ("max_seating", max_seating),
                    ("drivetrain", drivetrain), ("min_fuel_efficiency_city", min_fuel_efficiency_city),
                    ("min_cargo_space", min_cargo_space), ("has_sunroof", has_sunroof),
                    ("has_leather_seats", has_leather_seats), ("has_navigation", has_navigation),
                    ("has_backup_camera", has_backup_camera), ("min_safety_rating", min_safety_rating),
                ] if v is not None}
            }
    except sqlite3.Error as e:
        return {"count": 0, "cars": [], "status": "error", "error": str(e)}


@tool
def count_cars_only(
    min_price: Optional[int] = None,
    max_price: Optional[int] = None,
    fuel_type: Optional[str] = None,
    is_suv: Optional[bool] = None,
    min_seating: Optional[int] = None,
    max_seating: Optional[int] = None,
    drivetrain: Optional[str] = None,
    min_fuel_efficiency_city: Optional[float] = None,
    min_cargo_space: Optional[int] = None,
    has_sunroof: Optional[bool] = None,
    has_leather_seats: Optional[bool] = None,
    has_navigation: Optional[bool] = None,
    has_backup_camera: Optional[bool] = None,
    min_safety_rating: Optional[float] = None,
) -> dict:
    """Count how many cars match the criteria without returning full results."""
    conditions = []
    params = []

    if min_price is not None:
        conditions.append("price >= ?")
        params.append(min_price)
    if max_price is not None:
        conditions.append("price <= ?")
        params.append(max_price)
    if fuel_type is not None:
        conditions.append("fuel_type = ?")
        params.append(fuel_type)
    if is_suv is not None:
        conditions.append("is_suv = ?")
        params.append(1 if is_suv else 0)
    if min_seating is not None:
        conditions.append("seating_capacity >= ?")
        params.append(min_seating)
    if max_seating is not None:
        conditions.append("seating_capacity <= ?")
        params.append(max_seating)
    if drivetrain is not None:
        conditions.append("drivetrain = ?")
        params.append(drivetrain)
    if min_fuel_efficiency_city is not None:
        conditions.append("fuel_efficiency_city >= ?")
        params.append(min_fuel_efficiency_city)
    if min_cargo_space is not None:
        conditions.append("cargo_space >= ?")
        params.append(min_cargo_space)
    if has_sunroof is not None:
        conditions.append("has_sunroof = ?")
        params.append(1 if has_sunroof else 0)
    if has_leather_seats is not None:
        conditions.append("has_leather_seats = ?")
        params.append(1 if has_leather_seats else 0)
    if has_navigation is not None:
        conditions.append("has_navigation = ?")
        params.append(1 if has_navigation else 0)
    if has_backup_camera is not None:
        conditions.append("has_backup_camera = ?")
        params.append(1 if has_backup_camera else 0)
    if min_safety_rating is not None:
        conditions.append("safety_rating >= ?")
        params.append(min_safety_rating)

    where_clause = " AND ".join(conditions) if conditions else "1=1"
    query = f"SELECT COUNT(*) as count FROM cars WHERE {where_clause}"

    try:
        with sqlite3.connect(DB_PATH) as conn:
            cursor = conn.cursor()
            cursor.execute(query, params)
            count = cursor.fetchone()[0]

            return {
                "count": count,
                "status": "too_few" if count < MIN_RESULTS else "too_many" if count > MAX_RESULTS else "good"
            }
    except sqlite3.Error as e:
        return {"count": 0, "status": "error", "error": str(e)}


# ============================================================================
# LANGGRAPH STATE DEFINITION
# ============================================================================

class ConversationState(TypedDict):
    """State for the conversation graph with validation tracking"""
    messages: Annotated[list, add]
    search_params: Optional[dict]
    search_results: Optional[dict]
    iteration_count: int
    user_satisfied: bool
    requires_search: bool
    moderation_result: Optional[dict]  # NEW: Track moderation results
    quality_result: Optional[dict]     # NEW: Track quality validation results
    regenerate_count: int              # NEW: Track regeneration attempts


# ============================================================================
# LANGGRAPH NODES WITH VALIDATION
# ============================================================================

def moderate_input(state: ConversationState) -> ConversationState:
    """
    Node: Validate user input for harmful or inappropriate content.
    This runs BEFORE processing the user's message.
    """
    messages = state["messages"]

    # Get last user message
    user_messages = [m for m in messages if isinstance(m, HumanMessage)]
    if not user_messages:
        return {"moderation_result": {"is_safe": True, "should_block": False}}

    last_user_msg = user_messages[-1].content

    # Call moderation tool
    moderation_result = moderate_user_input.invoke({"user_message": last_user_msg})

    return {
        "moderation_result": moderation_result
    }


def gather_requirements(state: ConversationState) -> ConversationState:
    """
    Node: Gather requirements from user and determine search parameters.
    Only runs if moderation passed.
    """
    messages = state["messages"]

    system_prompt = f"""You are a car shopping advisor. Be concise and helpful.

{SCHEMA_DESCRIPTION}

CRITICAL - You must provide text explanation WITH your tool call:

When responding, ALWAYS include BOTH:

1. **Text content explaining your recommendation:**
   Example: "For driving in Paris, I'd recommend a compact sedan or hybrid with good fuel efficiency (25+ MPG) since city parking is tight and gas is expensive. Let me search our inventory..."

2. **Tool call to search:**
   Call search_cars with appropriate parameters based on user needs:

   Common mappings:
   - City driving/parking → max_price=35000, min_fuel_efficiency_city=25 (compact, efficient)
   - Family/kids → is_suv=True, min_seating=5, min_safety_rating=4.0
   - Fuel efficiency → min_fuel_efficiency_city=25 or fuel_type="Hybrid"/"Electric"
   - Budget conscious → max_price=30000
   - Luxury → has_leather_seats=True, has_navigation=True, min_price=35000
   - Outdoor/adventure → is_suv=True, drivetrain="AWD" or "4WD"
   - Long commute → min_fuel_efficiency_city=28, max_price=35000

IMPORTANT: Your response must have BOTH text content (recommendation) AND a tool_call (search). Never tool call without explanation text."""

    llm_with_tools = llm.bind_tools([search_cars, count_cars_only])
    response = llm_with_tools.invoke([SystemMessage(content=system_prompt)] + messages)

    new_state = {
        "messages": [response],
        "requires_search": bool(response.tool_calls),
        "iteration_count": state.get("iteration_count", 0)
    }

    if "satisfied" in response.content.lower() or "perfect" in response.content.lower():
        new_state["user_satisfied"] = True

    return new_state


def execute_search(state: ConversationState) -> ConversationState:
    """Node: Execute the search using tool calls from the LLM."""
    messages = state["messages"]
    last_message = messages[-1]

    if hasattr(last_message, 'tool_calls') and last_message.tool_calls:
        tool_node = ToolNode([search_cars, count_cars_only])
        result = tool_node.invoke(state)

        tool_messages = result["messages"]
        if tool_messages:
            tool_response = tool_messages[-1]
            if hasattr(tool_response, 'content'):
                import json
                try:
                    search_results = json.loads(tool_response.content)
                except:
                    search_results = {"error": "Failed to parse tool response"}
            else:
                search_results = {}
        else:
            search_results = {}

        return {
            "messages": tool_messages,
            "search_results": search_results,
            "iteration_count": state.get("iteration_count", 0) + 1
        }

    return {"iteration_count": state.get("iteration_count", 0)}


def present_results(state: ConversationState) -> ConversationState:
    """Node: Present search results to user and provide guidance."""
    search_results = state.get("search_results", {})
    messages = state["messages"]

    system_prompt = """Present search results very briefly. DO NOT re-explain recommendations.

STATUS RESPONSES:

**good (1-20 cars):**
"Found X vehicles matching these criteria. See the full list below."

**too_few (<1):**
"No matches. Try: broaden budget, relax features, or consider more fuel types."

**too_many (>20):**
"Over 20 matches. Let's narrow down - what's your priority: budget, efficiency, space, or safety?"

**error:**
"Search error. Please rephrase."

Keep it very short - 1 sentence. Don't describe the cars, just confirm results are ready."""

    context = f"\nSearch Results: {search_results}"

    response = llm.invoke([
        SystemMessage(content=system_prompt),
        *messages,
        HumanMessage(content=context)
    ])

    return {"messages": [response]}


def validate_response_quality(state: ConversationState) -> ConversationState:
    """
    Node: Validate the quality of the assistant's response.
    Checks relevance, role adherence, and helpfulness.
    """
    messages = state.get("messages", [])

    # Get last user and assistant messages
    user_messages = [m for m in messages if isinstance(m, HumanMessage)]
    ai_messages = [m for m in messages if isinstance(m, AIMessage)]

    if not user_messages or not ai_messages:
        return {"quality_result": {"should_regenerate": False, "is_relevant": True, "stays_in_role": True, "is_helpful": True}}

    last_user_msg = user_messages[-1].content if hasattr(user_messages[-1], 'content') else str(user_messages[-1])
    last_ai_msg = ai_messages[-1].content if hasattr(ai_messages[-1], 'content') else str(ai_messages[-1])

    # Call quality validation tool
    try:
        quality_result = validate_assistant_response.invoke({
            "user_message": last_user_msg,
            "assistant_response": last_ai_msg
        })
    except Exception as e:
        # If validation fails, assume response is acceptable
        quality_result = {
            "should_regenerate": False,
            "is_relevant": True,
            "stays_in_role": True,
            "is_helpful": True,
            "explanation": f"Validation error: {str(e)}"
        }

    return {
        "quality_result": quality_result
    }


# ============================================================================
# CONDITIONAL EDGE LOGIC
# ============================================================================

def should_moderate(state: ConversationState) -> Literal["moderate_input", "gather_requirements"]:
    """Decide if we need to moderate user input"""
    messages = state["messages"]

    # Check if last message is from user
    if messages and isinstance(messages[-1], HumanMessage):
        return "moderate_input"

    return "gather_requirements"


def check_moderation(state: ConversationState) -> Literal["gather_requirements", "end"]:
    """Check if moderation passed or blocked content"""
    moderation = state.get("moderation_result", {})

    if moderation.get("should_block", False):
        return "end"  # Block harmful content

    return "gather_requirements"


def should_continue(state: ConversationState) -> Literal["execute_search", "present_results", "validate_response", "end"]:
    """Determine next step in the graph"""
    if state.get("user_satisfied", False):
        return "end"

    if state.get("requires_search", False) and not state.get("search_results"):
        return "execute_search"

    if state.get("search_results"):
        return "present_results"

    # If we have an AI response, validate it
    messages = state.get("messages", [])
    if messages and isinstance(messages[-1], AIMessage):
        return "validate_response"

    return "end"


def check_quality(state: ConversationState) -> Literal["gather_requirements", "end"]:
    """Check if response needs regeneration"""
    quality = state.get("quality_result", {})
    regenerate_count = state.get("regenerate_count", 0)

    # Only regenerate once to avoid loops
    if quality.get("should_regenerate", False) and regenerate_count < 1:
        return "gather_requirements"

    return "end"


# ============================================================================
# BUILD LANGGRAPH WITH VALIDATION
# ============================================================================

def build_graph() -> StateGraph:
    """Build the LangGraph workflow with validation nodes"""

    workflow = StateGraph(ConversationState)

    # Add validation nodes
    workflow.add_node("moderate_input", moderate_input)
    workflow.add_node("validate_response", validate_response_quality)

    # Add core nodes
    workflow.add_node("gather_requirements", gather_requirements)
    workflow.add_node("execute_search", execute_search)
    workflow.add_node("present_results", present_results)

    # Set entry point - start with moderation
    workflow.set_entry_point("moderate_input")

    # Flow: moderate -> gather -> search -> present -> validate
    workflow.add_conditional_edges(
        "moderate_input",
        check_moderation,
        {
            "gather_requirements": "gather_requirements",
            "end": END
        }
    )

    workflow.add_conditional_edges(
        "gather_requirements",
        should_continue,
        {
            "execute_search": "execute_search",
            "present_results": "present_results",
            "validate_response": "validate_response",
            "end": END
        }
    )

    workflow.add_edge("execute_search", "present_results")
    workflow.add_edge("present_results", "validate_response")

    workflow.add_conditional_edges(
        "validate_response",
        check_quality,
        {
            "gather_requirements": "gather_requirements",
            "end": END
        }
    )

    return workflow.compile()


# ============================================================================
# HELPER FUNCTIONS
# ============================================================================

def format_car_display(car: dict) -> str:
    """Format a single car for display"""
    return f"""
{car['brand']} {car['model_name']} ({car['year']})
  Price: ${car['price']:,}
  Type: {'SUV' if car['is_suv'] else 'Sedan/Coupe'}
  Fuel: {car['fuel_type']}
  Seats: {car['seating_capacity']} | Cargo: {car['cargo_space']} cu ft
  Drivetrain: {car['drivetrain']} | Transmission: {car['transmission']}
  Features: {'Sunroof, ' if car['has_sunroof'] else ''}{'Leather, ' if car['has_leather_seats'] else ''}{'Navigation, ' if car['has_navigation'] else ''}{'Backup Camera' if car['has_backup_camera'] else ''}
"""


# ============================================================================
# MAIN CHATBOT LOOP
# ============================================================================

def main():
    """Main chatbot loop using LangGraph with validation"""
    print("=" * 60)
    print("Car Finder Chatbot (LangGraph + Validation Agents)")
    print("=" * 60)
    print("Features: Content Moderation + Response Quality Validation")
    print("Tell me what kind of car you're looking for!")
    print("Type 'quit' to exit.\n")

    app = build_graph()

    conversation_state = {
        "messages": [],
        "search_params": None,
        "search_results": None,
        "iteration_count": 0,
        "user_satisfied": False,
        "requires_search": False,
        "moderation_result": None,
        "quality_result": None,
        "regenerate_count": 0
    }

    while True:
        user_input = input("You: ").strip()

        if user_input.lower() in ['quit', 'exit', 'bye']:
            print("\nThanks for using Car Finder! Goodbye!")
            break

        if not user_input:
            continue

        # Add user message to state
        conversation_state["messages"].append(HumanMessage(content=user_input))

        try:
            result = app.invoke(conversation_state)
            conversation_state = result

            # Check if moderation blocked the message
            moderation = result.get("moderation_result", {})
            if moderation.get("should_block", False):
                print(f"\n[Content Warning] Your message was flagged as potentially {moderation.get('violation_type', 'inappropriate')}.")
                print(f"Reason: {moderation.get('explanation')}")
                print("Please keep the conversation professional and car-related.\n")
                # Remove the blocked message from history
                conversation_state["messages"] = conversation_state["messages"][:-1]
                continue

            # Display assistant's responses - show all new AI messages
            if result.get("messages"):
                # Get all AI messages that weren't in the previous state
                new_messages = []
                for msg in result["messages"]:
                    if isinstance(msg, AIMessage):
                        new_messages.append(msg)

                # Display all new AI messages with content
                for msg in new_messages:
                    if hasattr(msg, 'content') and msg.content:
                        print(f"\nAssistant: {msg.content}\n")

                # Show quality validation feedback (for debugging)
                quality = result.get("quality_result")
                if quality and not quality.get("is_relevant", True):
                    print(f"[Quality Warning] Response quality issues detected: {quality.get('explanation')}\n")

                # Display search results if available
                search_results = result.get("search_results")
                if search_results and search_results.get("status") == "good":
                    cars = search_results.get("cars", [])
                    if cars:
                        print("=" * 60)
                        print("MATCHING CARS:")
                        print("=" * 60)
                        for car in cars:
                            print(format_car_display(car))
                        print("=" * 60)
                        print()

            # Check if conversation should end
            if result.get("user_satisfied", False):
                print("\nThank you for using Car Finder! Have a great day!")
                break

        except Exception as e:
            print(f"\nError: {str(e)}")
            print("Let's try again. Please rephrase your request.\n")
            conversation_state["messages"] = conversation_state["messages"][:-1]


if __name__ == "__main__":
    main()