This tutorial explains how to implement an Agent-based Dynamic Slot Filling system. It covers the process of creating an intelligent conversational system that analyzes user requests to automatically collect necessary information and supplements missing information through dialogue.
The system can handle various tasks such as restaurant reservations, meeting scheduling, hotel bookings, and flight reservations, dynamically collecting and validating the required information for each task.
Features
Dynamic Slot Filling: Automatically identifies and collects necessary information by analyzing user requests
Multi-task Support: Handles various tasks including restaurant, meeting, hotel, and flight reservations
Conversational Information Collection: Supplements missing information through natural dialogue
Information Validation: Automatically validates the input information
Graph of Agent-Based Dynamic Slot Filling
This graph shows the workflow of an Agent-based Dynamic Slot Filling system:
Start → Classify
Analyzes user requests and classifies task type
Initialize Slots
Sets up required information fields
Extract Slots
Extracts necessary information from user messages
Identifies missing information
Generate Response
Requests additional information or completes task
Ends when all information is collected
The system iterates through conversation with the user until all necessary information is gathered.
Table of Contents
References
Environment Setup
[Note]
langchain-opentutorial is a package that provides a set of easy-to-use environment setup, useful functions and utilities for tutorials.
You can alternatively set API keys such as OPENAI_API_KEY in a .env file and load them.
[Note] This is not necessary if you've already set the required API keys in previous steps.
# Load API keys from .env file
from dotenv import load_dotenv
load_dotenv(override=True)
True
Constants and Prompt Template Definition
Constants is defined as below:
TASK_SLOTS : Defines the required information for each task type
TASK_RULES : Defines the rules for information collection for each task type
TASK_EXAMPLES : Provides examples of user requests for each task type
TASK_CLASSIFICATION_TEMPLATE : A prompt template for classifying the user's request into a task
SLOT_EXTRACTION_TEMPLATE : A prompt template for extracting information from the user's message
RESPONSE_TEMPLATE : A prompt template for generating a response to the user's message
Task Slots
Task Slots is a structure that defines the required information for each task type. Each task has its own unique set of slots, allowing systematic collection of necessary information.
TASK_SLOTS = {
"restaurant": {
"restaurant_address": "Restaurant Address/Location (City name)",
"number_of_people": "Number of People (numeric)",
"reservation_datetime": "Reservation Date and Time (YYYY/MM/DD HH:MM format)",
},
"meeting": {
"meeting_datetime": "Meeting Date and Time (YYYY/MM/DD HH:MM format)",
"platform": "Video Conference Platform (zoom/teams/google meet)",
"meeting_duration": "Meeting Duration (in minutes)",
},
"hotel": {
"hotel_location": "Hotel Location (City name)",
"check_in_date": "Check-in Date (YYYY/MM/DD)",
"check_out_date": "Check-out Date (YYYY/MM/DD)",
"room_type": "Room Type (single/double/suite)",
"number_of_guests": "Number of Guests (numeric)",
},
"flight": {
"departure_city": "Departure City",
"arrival_city": "Arrival City",
"departure_date": "Departure Date (YYYY/MM/DD HH:MM format)",
"return_date": "Return Date (YYYY/MM/DD HH:MM format)",
"passenger_count": "Number of Passengers (numeric)",
"seat_class": "Seat Class (economy/business/first)",
},
}
Task Rules
Defines the rules to follow when collecting information for each task type. These rules serve as guidelines that the conversational agent references when gathering information through user interactions.
Restaurant Reservation: Collect information about location, number of people, and reservation date/time
Hotel Booking: Collect information about location, check-in/out dates, room type, and number of guests
Meeting Room Reservation: Collect information about location, number of attendees, meeting date/time, and duration
Flight Reservation: Collect information about departure/arrival locations, departure/return dates, and number of passengers
TASK_RULES = {
"restaurant": """
1. Do not ask for information that has already been provided
2. First confirm the specific restaurant location (city name)
3. Accept only numeric values for the number of people
4. Accept reservation date and time in YYYY/MM/DD HH:MM format""",
"hotel": """
1. Do not ask for information that has already been provided
2. First confirm the hotel location (city name)
3. Accept check-in/check-out dates in YYYY/MM/DD format
4. Have users select room type from: single/double/suite
5. Accept only numeric values for the number of guests""",
"meeting": """
1. Do not ask for information that has already been provided
2. First confirm the meeting room location
3. Accept only numeric values for the number of attendees
4. Accept meeting date and time in YYYY/MM/DD HH:MM format
5. Accept meeting duration in hours (e.g., 1 hour, 2 hours)""",
"flight": """
1. Do not ask for information that has already been provided
2. First confirm departure and arrival locations
3. Accept departure and return dates in YYYY/MM/DD HH:MM format
4. Accept only numeric values for the number of passengers""",
}
TASK_EXAMPLES = {
"restaurant": "I'd like to make a dinner reservation for 4 people next Friday at 7 PM",
"hotel": "I want to book a suite room in Manhattan, New York from the 1st to the 3rd of next month",
"meeting": "I'm planning to have a one-hour meeting tomorrow at 2 PM in the Downtown conference room",
"flight": "I'd like to book 2 economy seats from LAX to New York at 10 AM on the 15th of next month",
}
Task Classification Template
Defines the prompt template for classifying the user's request into a task.
user_message : The user's message to be analyzed
task_type : The type of task selected by the agent
confidence : The confidence score of the task classification (0.0 ~ 1.0)
TASK_CLASSIFICATION_TEMPLATE = """Please analyze the user's message and select the appropriate task for reservation/booking.
Available task list:
- restaurant: Restaurant Reservation
- meeting: Video Conference Booking
- hotel: Hotel Reservation
- flight: Flight Booking
User message:
{user_message}
Please respond in the following format:
{{"task": "task_name", "confidence": 0.XX}}
confidence should be a value between 0.0 and 1.0, indicating the certainty of intent classification.
If the intent cannot be determined, set confidence to 0.
"""
Slot Extraction Template
Defines the prompt template for extracting information from the user's message.
task_type : The type of task for which information is being extracted
required_slots : The slots that need to be extracted
slots : The current state of the slots
messages : The conversation history
last_message : The last message from the user
Please follow these rules strictly:
Date and Time Conversion Rules:
All dates must be in YYYY/MM/DD HH:MM format8
If only "next week" is mentioned, ask for specific date and time (keep as null )
Convert to date only when day of week is specified (e.g., "next Monday")
If no time is specified, keep as null
Incomplete Date/Time Cases:
"Next week" only → null
"Evening" only → null
"Next Monday evening" → YYYY/MM/DD 19:00
"Tomorrow lunch" → YYYY/MM/DD 12:00
Numbers must be converted to numeric format (e.g., "four people" → 4 )
Use location names as is (e.g., "Manhattan", "New York")
Mark uncertain information as null
SLOT_EXTRACTION_TEMPLATE = """Please extract information related to {task_type} reservation from the following conversation.
Required Information:
{required_slots}
Current Slot Status:
{slots}
Conversation:
{messages}
Last Message:
{last_message}
Current Date: {current_date}
Please follow these rules strictly:
1. Date and Time Conversion Rules:
- All dates must be in YYYY/MM/DD HH:MM format
- If only "next week" is mentioned, ask for specific date and time (keep as null)
- Convert to date only when day of week is specified (e.g., "next Monday")
- If no time is specified, keep as null
2. Incomplete Date/Time Cases:
- "Next week" only → null
- "Evening" only → null
- "Next Monday evening" → YYYY/MM/DD 19:00
- "Tomorrow lunch" → YYYY/MM/DD 12:00
3. Numbers must be converted to numeric format (e.g., "four people" → "4")
4. Use location names as is (e.g., "Manhattan", "New York")
5. Mark uncertain information as null
Please respond with extracted information in the following JSON format:
{{"restaurant_address": "location or null",
"number_of_people": "number or null",
"reservation_datetime": "YYYY/MM/DD HH:MM or null"}}
"""
Response Template
Defines the prompt template for collecting missing information through natural dialogue.
task_type : The type of task for which information is being collected
required_slots : The slots that need to be collected
slots : The current state of the slots
messages : The conversation history
last_message : The last message from the user
Please follow these rules:
task_rules: The rules specific to the current task type
Respond in a natural, conversational manner
RESPONSE_TEMPLATE = """Continue the conversation in a friendly tone while collecting missing information.
Reservation Type: {task_type}
Required Information:
{required_slots}
Current Slot Status:
{slots}
Conversation History:
{messages}
Please follow these rules:
{task_rules}
Respond in a natural, conversational manner."""
from langchain_core.prompts import ChatPromptTemplate
slot_extraction_prompt = ChatPromptTemplate.from_template(SLOT_EXTRACTION_TEMPLATE)
response_prompt = ChatPromptTemplate.from_template(RESPONSE_TEMPLATE)
State Management
State management plays a crucial role in controlling the flow of the conversation and tracking necessary information.
Defines the SupervisorState class to manage the state of the conversational agent.
Inherits from TypedDict to define and track the state of the conversational agent.
This state management allows maintaining the conversation context with the user and sequentially collecting necessary information.
SupervisorState
messages : Manages conversation history
task_type : Tracks current task type
confidence : Task classification confidence score
slots : Stores collected information
current_slot : Currently processing slot
completed : Task completion status
stage : Current stage ('classify' or 'slot_filling')
from typing import TypedDict
from langgraph.graph import StateGraph, START, END
from langchain_core.messages import HumanMessage, AIMessage
from datetime import datetime
class SupervisorState(TypedDict):
"""Supervisor State for managing the state of the entire system"""
messages: list[HumanMessage | AIMessage]
task_type: str | None
confidence: float
slots: dict
current_slot: str | None
completed: bool
stage: str
Graph Construction
Uses LangGraph's StateGraph to construct the conversation flow.
Main Nodes
classify_task: Classifies the task type based on the user's message
Identifies reservation/booking intent from user message to select appropriate task
Proceeds to slot initialization or information extraction based on selected task
initialize_slots: Initializes slots for user message
Initializes required slots based on selected task
Stores initialized slot state in state variables
extract_slots: Extracts necessary information from user message
Uses LLM to extract structured information from natural language
Validates extracted information
Updates with new information while maintaining existing slot values
generate_response: Generates appropriate response based on current state
Branches response based on task classification confidence
Requests missing information
Generates reservation completion message
should_continue: Controls conversation flow by determining next step based on:
Checks user input waiting status
Branches based on task classification confidence
Checks if slot initialization is needed
Determines whether to continue information extraction
def initialize_slots(state: SupervisorState) -> SupervisorState:
"""Initializes slots based on task type."""
print("\n=== Initializing Slots ===")
if state["task_type"]:
state["slots"] = {
slot: "null" for slot in TASK_SLOTS[state["task_type"]].keys()
}
print(f"Initialized Slots: {state['slots']}")
return state
def extract_slots(state: SupervisorState) -> SupervisorState:
"""Extracts slot information from the conversation."""
print("\n=== Extracting Slot Information ===")
try:
llm = ChatOpenAI(temperature=0)
required_slots = "\n".join(
[f"- {k}: {v}" for k, v in TASK_SLOTS[state["task_type"]].items()]
)
messages_text = "\n".join(msg.content for msg in state["messages"])
last_message = state["messages"][-1].content
current_date = datetime.now().strftime("%Y/%m/%d")
chain = slot_extraction_prompt | llm
result = chain.invoke(
{
"task_type": state["task_type"],
"required_slots": required_slots,
"slots": json.dumps(state["slots"], ensure_ascii=False),
"messages": messages_text,
"last_message": last_message,
"current_date": current_date,
}
)
try:
new_slots = json.loads(result.content)
for slot, value in new_slots.items():
if (
value is not None
and str(value).lower() != "null"
and str(value).strip()
):
state["slots"][slot] = value
print("\n=== Current Slot Status ===")
print(f"Task Type: {state['task_type']}")
for slot, value in state["slots"].items():
print(f"{TASK_SLOTS[state['task_type']][slot]}: {value}")
print("=====================\n")
except json.JSONDecodeError:
print("Error parsing slot information.")
except Exception as e:
print(f"Error extracting slot information: {str(e)}")
return state
def generate_response(state: SupervisorState) -> SupervisorState:
"""Generates response based on current state."""
print("\n=== Generating Response ===")
response = ""
if state["stage"] == "classify" and state["confidence"] < 0.5:
response = "Sorry, I couldn't determine which type of reservation you're looking for.\n"
response += "The following reservations are possible:\n"
for task in TASK_SLOTS.keys():
response += f"- {task}\n"
response += (
"\nPlease specify the reservation you're looking for in more detail."
)
else:
empty_slots = []
for slot, value in state["slots"].items():
if value is None or str(value).lower() == "null" or not str(value).strip():
empty_slots.append(slot)
if empty_slots:
task_type = state["task_type"]
response = "Could you please provide the following information:\n"
for slot in empty_slots:
response += f"- {TASK_SLOTS[task_type][slot]}\n"
else:
response = (
"All information has been entered. I will complete the reservation."
)
state["completed"] = True
state["messages"].append(AIMessage(content=response))
print(f"\nAI: {response}")
return state
def should_continue(state: SupervisorState) -> str:
"""Determines the next step."""
print(f"\n=== Current Stage: {state['stage']} ===")
last_message = state["messages"][-1]
if isinstance(last_message, AIMessage):
print("Waiting for user input.")
return "generate_response"
if state["stage"] == "classify":
if state["confidence"] < 0.5:
print("Task classification confidence is low. Retrying.")
return "generate_response"
print("Task classification complete. Proceeding to slot initialization.")
return "initialize_slots"
if not state["slots"]:
print("Slots are not initialized. Initializing.")
return "initialize_slots"
all_slots_filled = all(
value is not None
and str(value) != "null"
and str(value).strip() != ""
and str(value) != "undefined"
for value in state["slots"].values()
)
if all_slots_filled:
print("All slots are filled. Completing reservation.")
state["completed"] = True
return "generate_response"
print("Additional information is needed. Continuing slot extraction.")
return "extract_slots"
Create Reservation Agent Graph
Creates the integrated reservation system Agent graph.
Uses StateGraph to define the conversation flow
Each node is connected to a function that performs a specific task
Conditional edges to control flow based on state
Execution Flow
Receive user input
Task type classification (classify_task)
Slot initialization (initialize_slots)
Information extraction (extract_slots)
Generate response (generate_response)
Repeat 2-5 if necessary
Complete reservation when all information is collected
This structure allows for collecting necessary information through natural conversation with the user and performing appropriate processing for each reservation type.
from IPython.display import Image, display
from langchain_core.runnables.graph import MermaidDrawMethod
# Visualize the compiled StateGraph as a Mermaid diagram
display(
Image(
reservation_agent.get_graph().draw_mermaid_png(
draw_method=MermaidDrawMethod.API,
)
)
)
Example Execution
Each example demonstrates how the system automatically extracts necessary information and completes the task through natural conversation.
Restaurant Reservation
User: I'd like to make a dinner reservation for 4 people next Friday at 7 PM
AI: Could you please specify the restaurant location (city name)?
User: New York
AI: All information has been entered. I will complete the reservation.
Hotel Booking
User: I want to book a suite room in Manhattan, New York from the 1st to the 3rd of next month
AI: Could you please specify the hotel location (city name)?
User: New York
AI: All information has been entered. I will complete the reservation.
Meeting Scheduling
User: I'm planning to have a one-hour meeting tomorrow at 2 PM in the Downtown conference room
AI: Could you please specify the video conference platform (zoom/teams/google meet)?
User: Zoom
AI: All information has been entered. I will complete the reservation.
Flight Booking
User: I'd like to book 2 economy seats from LAX to New York at 10 AM on the 15th of next month
AI: Could you please specify the departure city?
User: Los Angeles
AI: All information has been entered. I will complete the reservation.
def agent_chat():
"""Runs the interactive reservation system."""
reservation_agent = create_reservation_agent()
print("\n=== AI Reservation Assistant ===")
print("Feel free to discuss your desired reservation or request.")
print("\n📝 Example Phrases:")
print("├── Restaurant: " + TASK_EXAMPLES["restaurant"])
print("├── Hotel: " + TASK_EXAMPLES["hotel"])
print("├── Meeting Room: " + TASK_EXAMPLES["meeting"])
print("└── Flight: " + TASK_EXAMPLES["flight"])
print("\nTo exit, type 'quit' or 'exit'.\n")
messages_history = []
task_type = None
slots = {}
while True:
user_input = input("User: ").strip()
print(f"\nUser: {user_input}") # Add this line to echo user input
if user_input.lower() in ["quit", "exit"]:
print("Exiting reservation system.")
break
if not user_input:
continue
try:
messages_history.append(HumanMessage(content=user_input))
if not task_type:
state = {
"messages": messages_history,
"task_type": None,
"confidence": 0.0,
"slots": {},
"current_slot": None,
"completed": False,
"stage": "classify",
}
state = reservation_agent.nodes["classify"].invoke(state)
if state["confidence"] >= 0.5:
task_type = state["task_type"]
state = reservation_agent.nodes["initialize_slots"].invoke(state)
slots = state["slots"]
else:
print(
"Sorry, I couldn't determine which type of reservation you're looking for."
)
continue
state = {
"messages": messages_history,
"task_type": task_type,
"confidence": 1.0,
"slots": slots,
"current_slot": None,
"completed": False,
"stage": "slot_filling",
}
state = reservation_agent.nodes["extract_slots"].invoke(state)
slots = state["slots"]
state = reservation_agent.nodes["generate_response"].invoke(state)
if isinstance(state["messages"][-1], AIMessage):
messages_history.append(state["messages"][-1])
all_slots_filled = all(
value is not None and str(value) != "null" and str(value).strip()
for value in slots.values()
)
if all_slots_filled:
print("\n=== 📝 Conversation Summary ===")
for msg in messages_history:
prefix = "User: " if isinstance(msg, HumanMessage) else "AI: "
print(f"{prefix}{msg.content}")
print("\n=== ✨ Reservation Complete! ✨ ===")
print(f"Reservation Type: {task_type}")
for slot, value in slots.items():
print(f"{TASK_SLOTS[task_type][slot]}: {value}")
print("\nTo start a new reservation, feel free to discuss.")
print("\nTo exit, type 'quit' or 'exit'.\n")
messages_history = []
task_type = None
slots = {}
except Exception as e:
print(f"An error occurred: {str(e)}")
print("Please try again.")
continue
agent_chat()
=== AI Reservation Assistant ===
Feel free to discuss your desired reservation or request.
📝 Example Phrases:
├── Restaurant: I'd like to make a dinner reservation for 4 people next Friday at 7 PM
├── Hotel: I want to book a suite room in Manhattan, New York from the 1st to the 3rd of next month
├── Meeting Room: I'm planning to have a one-hour meeting tomorrow at 2 PM in the Downtown conference room
└── Flight: I'd like to book 2 economy seats from LAX to New York at 10 AM on the 15th of next month
To exit, type 'quit' or 'exit'.
User: I want to book a suite room in Manhattan, New York from the 1st to the 3rd of next month
=== Task Classification ===
Classified Task: hotel
Confidence: 1.00
=== Initializing Slots ===
Initialized Slots: {'hotel_location': 'null', 'check_in_date': 'null', 'check_out_date': 'null', 'room_type': 'null', 'number_of_guests': 'null'}
=== Extracting Slot Information ===
=== Current Slot Status ===
Task Type: hotel
Hotel Location (City name): New York
Check-in Date (YYYY/MM/DD): 2025/03/01
Check-out Date (YYYY/MM/DD): 2025/03/03
Room Type (single/double/suite): suite
Number of Guests (numeric): null
=====================
=== Generating Response ===
AI: Could you please provide the following information:
- Number of Guests (numeric)
User: 2
=== Extracting Slot Information ===
=== Current Slot Status ===
Task Type: hotel
Hotel Location (City name): New York
Check-in Date (YYYY/MM/DD): 2025/03/01
Check-out Date (YYYY/MM/DD): 2025/03/03
Room Type (single/double/suite): suite
Number of Guests (numeric): 2
=====================
=== Generating Response ===
AI: All information has been entered. I will complete the reservation.
=== 📝 Conversation Summary ===
User: I want to book a suite room in Manhattan, New York from the 1st to the 3rd of next month
AI: Could you please provide the following information:
- Number of Guests (numeric)
AI: Could you please provide the following information:
- Number of Guests (numeric)
User: 2
AI: All information has been entered. I will complete the reservation.
AI: All information has been entered. I will complete the reservation.
=== ✨ Reservation Complete! ✨ ===
Reservation Type: hotel
Hotel Location (City name): New York
Check-in Date (YYYY/MM/DD): 2025/03/01
Check-out Date (YYYY/MM/DD): 2025/03/03
Room Type (single/double/suite): suite
Number of Guests (numeric): 2
To start a new reservation, feel free to discuss.
To exit, type 'quit' or 'exit'.
User: I'm planning to have a one-hour meeting tomorrow at 2 PM in the Downtown conference room
=== Task Classification ===
Classified Task: meeting
Confidence: 1.00
=== Initializing Slots ===
Initialized Slots: {'meeting_datetime': 'null', 'platform': 'null', 'meeting_duration': 'null'}
=== Extracting Slot Information ===
=== Current Slot Status ===
Task Type: meeting
Meeting Date and Time (YYYY/MM/DD HH:MM format): 2025/02/06 14:00
Video Conference Platform (zoom/teams/google meet): null
Meeting Duration (in minutes): 60
=====================
=== Generating Response ===
AI: Could you please provide the following information:
- Video Conference Platform (zoom/teams/google meet)
User: zoom
=== Extracting Slot Information ===
=== Current Slot Status ===
Task Type: meeting
Meeting Date and Time (YYYY/MM/DD HH:MM format): 2025/02/06 14:00
Video Conference Platform (zoom/teams/google meet): zoom
Meeting Duration (in minutes): 60
=====================
=== Generating Response ===
AI: All information has been entered. I will complete the reservation.
=== 📝 Conversation Summary ===
User: I'm planning to have a one-hour meeting tomorrow at 2 PM in the Downtown conference room
AI: Could you please provide the following information:
- Video Conference Platform (zoom/teams/google meet)
AI: Could you please provide the following information:
- Video Conference Platform (zoom/teams/google meet)
User: zoom
AI: All information has been entered. I will complete the reservation.
AI: All information has been entered. I will complete the reservation.
=== ✨ Reservation Complete! ✨ ===
Reservation Type: meeting
Meeting Date and Time (YYYY/MM/DD HH:MM format): 2025/02/06 14:00
Video Conference Platform (zoom/teams/google meet): zoom
Meeting Duration (in minutes): 60
To start a new reservation, feel free to discuss.
To exit, type 'quit' or 'exit'.
User: I'd like to book 2 economy seats from LAX to New York at 10 AM on the 15th of next month
=== Task Classification ===
Classified Task: flight
Confidence: 1.00
=== Initializing Slots ===
Initialized Slots: {'departure_city': 'null', 'arrival_city': 'null', 'departure_date': 'null', 'return_date': 'null', 'passenger_count': 'null', 'seat_class': 'null'}
=== Extracting Slot Information ===
=== Current Slot Status ===
Task Type: flight
Departure City: LAX
Arrival City: New York
Departure Date (YYYY/MM/DD HH:MM format): 2025/03/15 10:00
Return Date (YYYY/MM/DD HH:MM format): null
Number of Passengers (numeric): 2
Seat Class (economy/business/first): economy
=====================
=== Generating Response ===
AI: Could you please provide the following information:
- Return Date (YYYY/MM/DD HH:MM format)
User: 2025/05/15 10:00
=== Extracting Slot Information ===
=== Current Slot Status ===
Task Type: flight
Departure City: LAX
Arrival City: New York
Departure Date (YYYY/MM/DD HH:MM format): 2025/03/15 10:00
Return Date (YYYY/MM/DD HH:MM format): 2025/05/15 10:00
Number of Passengers (numeric): 2
Seat Class (economy/business/first): economy
=====================
=== Generating Response ===
AI: All information has been entered. I will complete the reservation.
=== 📝 Conversation Summary ===
User: I'd like to book 2 economy seats from LAX to New York at 10 AM on the 15th of next month
AI: Could you please provide the following information:
- Return Date (YYYY/MM/DD HH:MM format)
AI: Could you please provide the following information:
- Return Date (YYYY/MM/DD HH:MM format)
User: 2025/05/15 10:00
AI: All information has been entered. I will complete the reservation.
AI: All information has been entered. I will complete the reservation.
=== ✨ Reservation Complete! ✨ ===
Reservation Type: flight
Departure City: LAX
Arrival City: New York
Departure Date (YYYY/MM/DD HH:MM format): 2025/03/15 10:00
Return Date (YYYY/MM/DD HH:MM format): 2025/05/15 10:00
Number of Passengers (numeric): 2
Seat Class (economy/business/first): economy
To start a new reservation, feel free to discuss.
To exit, type 'quit' or 'exit'.
User: exit
Exiting reservation system.
Set up the environment. You may refer to for more details.
