Home / Articles / Practical notes: Agentic AI Project: Build a Customer Service Chatbot for a

This article is published in English.

Practical notes: Agentic AI Project: Build a Customer Service Chatbot for a

Operable walkthrough of Practical notes: Agentic AI Project: Build a Customer Service Chatbot for a: contracts, checks, and drop-in code slots for teams shipping this pattern.

5381 words

The following notes reconstruct a practical path around “Agentic AI Project: Build a Customer Service Chatbot for a Clinic”. Emphasis stays on contracts, checks, and drop-in code placeholders rather than motivational framing. When working through the Overview stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

Introduction

The Introduction stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Keep graph state flat and typed. Nested blobs hide which node wrote which field and break resume after interrupts.

Problem Statement

The Problem Statement stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Keep graph state flat and typed. Nested blobs hide which node wrote which field and break resume after interrupts.

Solution

The Solution stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Keep graph state flat and typed. Nested blobs hide which node wrote which field and break resume after interrupts. The Solution stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

Setup

For the Setup stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.

Mac / Linux / Windows

For the Mac Linux Windows stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.

python -m venv .venv
source .venv/bin/activate
.venv\Scripts\Activate.ps1
.venv\Scripts\activate
(.venv) your-folder-name %
streamlit>=1.50.0
python-dotenv==1.0.0
pydantic==2.12.5
pandas==2.3.3
python-dateutil==2.8.2
langgraph>=1.0.7
openai>=2.16.0
pygraphviz==1.14
cd clinic-agent
pip install -r requirements.txt
OPENAI_API_KEY=your_openai_api_key_here

Part 1: Database Setup (data/db.py)

For the Part 1 Database Setup stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness. For the Part 1 Database Setup stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

1. Doctors Table

When working through the 1 Doctors Table stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Checkpoint after expensive steps. Resume should not re-bill the same LLM call when an operator retries a later node.

2. Customers Table

When working through the 2 Customers Table stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Checkpoint after expensive steps. Resume should not re-bill the same LLM call when an operator retries a later node.

3. Booking Table

When working through the 3 Booking Table stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Checkpoint after expensive steps. Resume should not re-bill the same LLM call when an operator retries a later node. When working through the 3 Booking Table stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

4. Creating the Database

The 4 Creating the Database stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Keep graph state flat and typed. Nested blobs hide which node wrote which field and break resume after interrupts.

# data/db.py - Database initialization and operations

import sqlite3
import os
from datetime import datetime, timedelta

DB_PATH = os.path.join(os.path.dirname(__file__), "clinic.db")

def get_connection():
    """Get a database connection."""
    return sqlite3.connect(DB_PATH)

def init_db():
    """Initialize the database with tables and sample data."""
    conn = get_connection()
    cursor = conn.cursor()

    # Doctors table
    cursor.execute("""
        CREATE TABLE IF NOT EXISTS doctors (
            doctor_id TEXT PRIMARY KEY,
            doctor_name TEXT NOT NULL,
            speciality TEXT NOT NULL,
            office_timing TEXT NOT NULL
        )
    """)

    # Customers table
    cursor.execute("""
        CREATE TABLE IF NOT EXISTS customers (
            customer_id TEXT PRIMARY KEY,
            name TEXT NOT NULL,
            phone TEXT NOT NULL
        )
    """)

    # Bookings table
    cursor.execute("""
        CREATE TABLE IF NOT EXISTS bookings (
            booking_id TEXT PRIMARY KEY,
            doctor_id TEXT NOT NULL,
            customer_id TEXT NOT NULL,
            appointment_date TEXT NOT NULL,
            appointment_time TEXT NOT NULL,
            status TEXT NOT NULL,
            FOREIGN KEY (doctor_id) REFERENCES doctors (doctor_id),
            FOREIGN KEY (customer_id) REFERENCES customers (customer_id)
        )
    """)

    # Insert sample doctors
    doctors = [
        ("D1", "Dr. Anil Sharma", "General Physician", "10:00-14:00"),
        ("D2", "Dr. Neha Verma", "Dermatologist", "11:00-16:00"),
        ("D3", "Dr. Rohit Mehta", "Orthopedic", "09:00-13:00"),
        ("D4", "Dr. Kavita Rao", "Pediatrician", "10:00-15:00"),
        ("D5", "Dr. Sanjay Iyer", "ENT Specialist", "12:00-17:00"),
    ]

    for doctor in doctors:
        cursor.execute(
            "INSERT OR IGNORE INTO doctors (doctor_id, doctor_name, speciality, office_timing) VALUES (?, ?, ?, ?)",
            doctor
        )

    conn.commit()
    conn.close()

if __name__ == "__main__":
    print("Initializing database...")
    init_db()
    print("Database initialized successfully.")
cd clinic-agent
python data/db.py

Part 2: Service Layer (Tools for the Agent)

The Part 2 Service Layer stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve.

Doctor Service (services/doctor_service.py)

The Doctor Service services doctorservice stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Keep graph state flat and typed. Nested blobs hide which node wrote which field and break resume after interrupts. The Doctor Service services doctorservice stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

# services/doctor_service.py - Doctor operations

from data.db import get_all_doctors, get_doctor_by_speciality, get_doctor_by_id

def get_specialities_list():
    """Get list of all specialities."""
    doctors = get_all_doctors()
    # Return unique specialities
    return list(dict.fromkeys([doc[2] for doc in doctors]))

def get_doctor_info(speciality):
    """Get doctor information by speciality."""
    doctor = get_doctor_by_speciality(speciality)
    if doctor:
        return {
            "doctor_id": doctor[0],
            "doctor_name": doctor[1],
            "speciality": doctor[2],
            "office_timing": doctor[3]
        }
    return None

def generate_time_slots(office_timing):
    """Generate hourly time slots from office timing string.

    Args:
        office_timing: String like "11:00-16:00"

    Returns:
        List of time slots like ["11:00 AM", "12:00 PM", ...]
    """
    start_time, end_time = office_timing.split("-")
    start_hour = int(start_time.split(":")[0])
    end_hour = int(end_time.split(":")[0])

    slots = []
    for hour in range(start_hour, end_hour):
        if hour < 12:
            suffix = "AM"
            display_hour = hour if hour > 0 else 12
        elif hour == 12:
            suffix = "PM"
            display_hour = 12
        else:
            suffix = "PM"
            display_hour = hour - 12
        slots.append(f"{display_hour}:00 {suffix}")

    return slots

def parse_time_slot(slot_str):
    """Parse time slot string to 24-hour format.

    Args:
        slot_str: String like "1:00 PM"

    Returns:
        String like "13:00"
    """
    time_part, suffix = slot_str.split(" ")
    hour, minute = time_part.split(":")
    hour = int(hour)

    if suffix == "PM" and hour != 12:
        hour += 12
    elif suffix == "AM" and hour == 12:
        hour = 0

    return f"{hour:02d}:{minute}"

Booking Service (services/booking_service.py)

For the Booking Service services bookingservice stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.

# services/booking_service.py - Booking operations

import uuid
from datetime import datetime
from data.db import (
    create_customer,
    create_booking,
    get_customer_by_phone,
    get_bookings_by_doctor_and_date,
    get_booking_by_id
)
from services.doctor_service import parse_time_slot

def get_or_create_customer(name, phone):
    """Get existing customer or create new one."""
    customer = get_customer_by_phone(phone)
    if customer:
        return customer[0]  # Return customer_id

    customer_id = f"CUST-{uuid.uuid4().hex[:6].upper()}"
    create_customer(customer_id, name, phone)
    return customer_id

def get_available_slots(doctor_id, office_timing):
    """Get available time slots for a doctor for today.

    Args:
        doctor_id: Doctor ID
        office_timing: Office timing string like "11:00-16:00"

    Returns:
        List of available time slots
    """
    from services.doctor_service import generate_time_slots

    today = datetime.now().strftime("%Y-%m-%d")
    all_slots = generate_time_slots(office_timing)

    # Get booked slots
    booked_times = get_bookings_by_doctor_and_date(doctor_id, today)

    # Filter out booked slots
    available = []
    for slot in all_slots:
        slot_24h = parse_time_slot(slot)
        if slot_24h not in booked_times:
            available.append(slot)

    return available

def confirm_booking(doctor_id, customer_name, customer_phone, time_slot, appointment_date=None):
    """Confirm a booking.

    Args:
        doctor_id: Doctor ID
        customer_name: Customer name
        customer_phone: Customer phone
        time_slot: Time slot like "1:00 PM"
        appointment_date: Optional date in YYYY-MM-DD format. Defaults to today.

    Returns:
        Booking ID
    """
    # Get or create customer
    customer_id = get_or_create_customer(customer_name, customer_phone)

    # Generate booking ID
    booking_id = f"BKG-{uuid.uuid4().hex[:6].upper()}"

    # Format appointment time
    if not appointment_date:
        appointment_date = datetime.now().strftime("%Y-%m-%d")
    appointment_time = parse_time_slot(time_slot)

    # Create booking
    create_booking(booking_id, doctor_id, customer_id, appointment_date, appointment_time)

    return booking_id

Testing the service

For the Testing the service stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.

# test/test_service.py - to test the services created

from pathlib import Path
import sys

# Allow running this file directly: `python test/test_service.py`.
PROJECT_ROOT = Path(__file__).resolve().parents[1]
if str(PROJECT_ROOT) not in sys.path:
    sys.path.insert(0, str(PROJECT_ROOT))

from services.doctor_service import get_specialities_list, generate_time_slots
from services.booking_service import confirm_booking

# Get all specialities
specialities = get_specialities_list()
print("Available specialities:", specialities)

# Generate time slots for a doctor (11:00 AM - 4:00 PM)
slots = generate_time_slots("11:00-16:00")
print("Available slots:", slots)

# Confirm a booking
booking_id = confirm_booking(
    doctor_id="D1",
    customer_name="John Doe",
    customer_phone="9876543210",
    time_slot="2:00 PM"
)
print(f"Booking confirmed: {booking_id}")
cd clinic-agent
python test_service.py
(.venv) (base) my-mac clinic-agent % python test_service.py
Available specialities: ['General Physician', 'Dermatologist', 'Orthopedic', 'Pediatrician', 'ENT Specialist']
Available slots: ['11:00 AM', '12:00 PM', '1:00 PM', '2:00 PM', '3:00 PM']
Booking confirmed: BKG-C4F60A

Part 3: Agentic Layer

For the Part 3 Agentic Layer stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness. For the Part 3 Agentic Layer stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

Booking State:

When working through the Booking State stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Checkpoint after expensive steps. Resume should not re-bill the same LLM call when an operator retries a later node.

# agents/booking_agent.py - LangGraph agent implementation

from typing import TypedDict, Annotated, List, Optional
from langgraph.graph import StateGraph, END
from openai import OpenAI
import os
from dotenv import load_dotenv

load_dotenv()
client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))

class BookingState(TypedDict):
    """State for the booking conversation."""
    messages: List[dict]                    # Chat history
    stage: str                               # greeting, select_speciality, select_doctor, etc.
    selected_speciality: Optional[str]      # Chosen medical specialty
    selected_doctor: Optional[dict]         # Selected doctor details
    selected_date: Optional[str]            # Appointment date
    selected_slot: Optional[str]            # Time slot
    customer_name: Optional[str]            # Customer name
    customer_phone: Optional[str]           # Customer phone
    booking_id: Optional[str]               # Confirmation ID
    available_options: List[str]            # UI options

def create_initial_state():
    """Create initial state for the conversation."""
    return {
        "messages": [],
        "stage": "greeting",
        "selected_speciality": None,
        "selected_doctor": None,
        "selected_date": None,
        "selected_slot": None,
        "customer_name": None,
        "customer_phone": None,
        "booking_id": None,
        "available_options": []
    }

LLM Helper Function

When working through the LLM Helper Function stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Cache stable system instructions and tool schemas. Re-sending identical preamble is a common source of burn.

# agents/booking_agent.py - LangGraph agent implementation
def call_llm(
    system_prompt: str,
    user_prompt: str,
    *,
    model: str = "gpt-4o-mini",
    temperature: float = 0,
    max_tokens: int = 50,
) -> str:
    """
    Centralized helper for all LLM calls.
    Returns the assistant's response
    """
    try:
        response = client.chat.completions.create(
            model=model,
            messages=[
                {"role": "system", "content": system_prompt},
                {"role": "user", "content": user_prompt},
            ],
            temperature=temperature,
            max_tokens=max_tokens,
        )
        return response
    except Exception as e:
        print(f"LLM call error: {e}")
        return ""

Agent Nodes

When working through the Agent Nodes stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Checkpoint after expensive steps. Resume should not re-bill the same LLM call when an operator retries a later node. When working through the Agent Nodes stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

Building the Graph

The Building the Graph stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Keep graph state flat and typed. Nested blobs hide which node wrote which field and break resume after interrupts.

# agents/booking_agent.py - LangGraph agent implementation
from langgraph.checkpoint.memory import MemorySaver

def build_booking_graph():
    """Build the LangGraph workflow."""
    workflow = StateGraph(BookingState)

    # Add all nodes
    workflow.add_node("greeting", greeting_node)
    workflow.add_node("select_speciality", select_speciality_node)
    workflow.add_node("select_doctor", select_doctor_node)
    workflow.add_node("select_date", select_date_node)
    workflow.add_node("select_slot", select_slot_node)
    workflow.add_node("confirm", confirm_node)
    workflow.add_node("collect_details", collect_details_node)
    workflow.add_node("completed", completed_node)
    workflow.add_node("cancelled", cancelled_node)

    # Set entry point
    workflow.set_entry_point("greeting")

    # Add conditional edges based on routing
    workflow.add_conditional_edges(
        "greeting",
        llm_router,
        {
            "greeting": "greeting",
            "select_speciality": "select_speciality",
            "cancelled": "cancelled"
        }
    )

    # Similar conditional edges for other nodes...

    # Final edges to END
    workflow.add_edge("completed", END)
    workflow.add_edge("cancelled", END)

    # Compile with checkpointer for session management
    return workflow.compile(checkpointer=MemorySaver())

# Create the compiled graph
booking_graph = build_booking_graph()
#agents/save_langgraph_flow.py
"""Save the clinic booking LangGraph flow to png format in this folder."""


from pathlib import Path
import sys


# Ensure imports work whether the script is run from project root or this folder.
AGENTS_DIR = Path(__file__).resolve().parent
PROJECT_ROOT = AGENTS_DIR.parent
if str(PROJECT_ROOT) not in sys.path:
    sys.path.insert(0, str(PROJECT_ROOT))

from agents.booking_agent import booking_graph  # noqa: E402


def save_graph_files() -> None:
    """Export graph as PNG."""
    graph = booking_graph.get_graph()

    png_path = AGENTS_DIR / "langgraph_flow.png"
    png_data = graph.draw_mermaid_png()
    png_path.write_bytes(png_data)
    print(f"Saved PNG flow to: {png_path}")


if __name__ == "__main__":
    save_graph_files()
cd clinic-agent
python save_langgraph_flow.py

Process Message:

The Process Message stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Keep graph state flat and typed. Nested blobs hide which node wrote which field and break resume after interrupts.

# agents/booking_agent.py - LangGraph agent implementation
def process_message(state: BookingState, user_message: str, thread_id: str = "default_session") -> BookingState:
    """Process a user message through the booking graph."""
    config = {"configurable": {"thread_id": thread_id}}

    # Check if the graph is currently interrupted
    current_state = booking_graph.get_state(config)

    if current_state.tasks and current_state.tasks[0].interrupts:
        # Resume the graph with the user's message
        result = booking_graph.invoke(Command(resume=user_message), config=config)
    else:
        # No interrupt, so start/continue normally
        # Add user message to state (unless it's an initial trigger)
        if user_message.lower() != "hi" or state["messages"]:
            # Avoid duplicate user messages if already added
            if not state["messages"] or state["messages"][-1].get("content") != user_message:
                state["messages"].append({
                    "role": "user",
                    "content": user_message
                })
        # Run the graph
        result = booking_graph.invoke(state, config=config)

    # Update available_options and ensure message is in history
    snapshot = booking_graph.get_state(config)
    if snapshot.tasks and snapshot.tasks[0].interrupts:
        interrupt_value = snapshot.tasks[0].interrupts[0].value

        # Handle both dict and string interrupt values
        msg_content = ""
        options = []
        if isinstance(interrupt_value, dict):
            msg_content = interrupt_value.get("content", "")
            options = interrupt_value.get("available_options", [])
        else:
            msg_content = str(interrupt_value)

        # Ensure the interrupt message is in the chat history
        if msg_content:
            # Check if it was already added by the node
            last_msg_content = result["messages"][-1].get("content", "") if result["messages"] else ""
            if last_msg_content != msg_content:
                result["messages"].append({
                    "role": "assistant",
                    "content": msg_content,
                    "options": options
                })
            else:
                # If already added, just update it with options if missing
                result["messages"][-1]["options"] = options

        result["available_options"] = options
    else:
        # If not interrupted, use whatever set in state, or default to empty
        if "available_options" not in result:
            result["available_options"] = []

    return result

Testing the Agent

The Testing the Agent stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Keep graph state flat and typed. Nested blobs hide which node wrote which field and break resume after interrupts. The Testing the Agent stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

# test/test_agent.py - test agent
# Initialize state

from pathlib import Path
import sys

# Allow running this file directly: `python test/test_agent.py`.
PROJECT_ROOT = Path(__file__).resolve().parents[1]
if str(PROJECT_ROOT) not in sys.path:
 sys.path.insert(0, str(PROJECT_ROOT))

from agents.booking_agent import create_initial_state, process_message


state = create_initial_state()

# Process messages
state = process_message(state, "Hi", thread_id="session_1")
print(state["messages"][-1]["content"])

state = process_message(state, "I want to book", thread_id="session_1")
print(state["available_options"])
cd clinic-agent
python test/test_agent.py

Part 4: Streamlit UI

For the Part 4 Streamlit UI stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.

# ui/chat_ui.py - Streamlit chatbot interface

"""Streamlit UI for the clinic booking chatbot."""

import streamlit as st
from agents.booking_agent import create_initial_state, process_message
from data.db import init_db


def initialize_session():
    """Initialize session state."""
    if "state" not in st.session_state:
        st.session_state.state = create_initial_state()
    if "initialized" not in st.session_state:
        st.session_state.initialized = False
    if "session_id" not in st.session_state:
        import uuid
        st.session_state.session_id = str(uuid.uuid4())


def display_chat_history():
    """Display the chat history with persistent options and styling."""
    messages = st.session_state.state.get("messages", [])
    for i, message in enumerate(messages):
        if message["role"] == "assistant":
            with st.chat_message("assistant"):
                st.markdown(message["content"])

                # Show options if they exist
                options = message.get("options", [])
                if options:
                    # If this is the last message in history, show as clickable buttons
                    if i == len(messages) - 1 and st.session_state.state["stage"] not in ["completed", "cancelled"]:
                        st.markdown("---")
                        # Create columns for buttons
                        cols = st.columns(min(len(options), 3))
                        for idx, option in enumerate(options):
                            col_idx = idx % 3
                            with cols[col_idx]:
                                if st.button(option, key=f"btn_{i}_{idx}", use_container_width=True):
                                    handle_user_input(option)
                    else:
                        # For older messages, show options as pills/text to keep history
                        options_str = "  ".join([f"`{opt}`" for opt in options])
                        st.markdown(f"**Available options:** {options_str}")
        else:
            with st.chat_message("user"):
                st.markdown(message["content"])


def handle_user_input(user_input: str):
    """Handle user input and process through agent."""
    # Process the message
    st.session_state.state = process_message(
        st.session_state.state,
        user_input,
        thread_id=st.session_state.session_id
    )

    # Rerun to update UI
    st.rerun()


def run_chat_ui():
    """Run the chat UI."""
    # Page config
    st.set_page_config(
        page_title="CarePlus Clinic - Book Appointment",
        page_icon="🏥",
        layout="centered"
    )

    # Custom CSS for distinction between messages
    st.markdown("""
        <style>
        [data-testid="stChatMessageUser"] {
            flex-direction: row-reverse;
            text-align: right;
            background-color: #e0f2f1;
            border-radius: 15px 15px 0px 15px;
        }
        [data-testid="stChatMessageAssistant"] {
            background-color: #f5f5f5;
            border-radius: 15px 15px 15px 0px;
        }
        </style>
    """, unsafe_allow_html=True)

    # Initialize database
    init_db()

    # Initialize session
    initialize_session()

    # Header
    st.title("🏥 CarePlus Clinic")
    st.markdown("*Book your doctor appointment easily*")
    st.markdown("---")

    # Send initial greeting if not initialized
    if not st.session_state.initialized:
        st.session_state.state = process_message(
            st.session_state.state,
            "Hi",
            thread_id=st.session_state.session_id
        )
        st.session_state.initialized = True
        st.rerun()

    # Display chat history
    display_chat_history()

    # Chat input (only show if not completed)
    if st.session_state.state["stage"] not in ["completed", "cancelled"]:
        if prompt := st.chat_input("Type your message here..."):
            handle_user_input(prompt)
    else:
        # Show restart button after completion
        st.markdown("---")
        if st.button("🔄 Start New Booking", use_container_width=True):
            st.session_state.state = create_initial_state()
            st.session_state.initialized = False
            st.rerun()

Part 5: Application Entry Point

For the Part 5 Application Entry stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.

# app.py - Main entry point for the application

from ui.chat_ui import run_chat_ui

if __name__ == "__main__":
    run_chat_ui()

Running the Chatbot: Two Approaches

For the Running the Chatbot Two stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.

Approach 1: Streamlit Web Application

For the Approach 1 Streamlit Web stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.

streamlit run app.py

Approach 2: Jupyter Notebook

For the Approach 2 Jupyter Notebook stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.

# clinic-agent.ipynb
from services.doctor_service import get_specialities_list, get_doctor_info, generate_time_slots
from services.booking_service import confirm_booking
from agents.booking_agent import (
    BookingState,
    create_initial_state,
    build_booking_graph,
    process_message
)
# clinic-agent.ipynb
# Initialize the booking graph
booking_graph = build_booking_graph()

## Visualize the booking graph structure
from IPython.display import Image, display

png_bytes = booking_graph.get_graph().draw_mermaid_png()
display(Image(png_bytes))

Run the Chatbot Session:

For the Run the Chatbot Session stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.

# clinic-agent.ipynb
from langgraph.types import Command

def run_booking_session(graph, thread_id="notebook_session", reset=False):
    config = {"configurable": {"thread_id": thread_id}}

    # 1. Start or Reset logic
    current_state = graph.get_state(config)
    if reset or not current_state.values:
        print(f"--- {'🔄 Resetting' if reset else '🆕 Initializing'} Session ---")
        # Using invoke() here kicks off the 'greeting' node immediately
        graph.invoke(create_initial_state(), config=config)

    print("---⚕⚕ Starting CarePlus Booking Session ---")

    last_displayed_message_idx = -1  # Track which messages have been displayed

    while True:
        state = graph.get_state(config)

        # Display any new assistant messages that haven't been shown yet
        # (This handles guardrail/off-topic responses)
        if state.values and state.values.get('messages'):
            messages = state.values['messages']
            for idx in range(last_displayed_message_idx + 1, len(messages)):
                msg = messages[idx]
                if msg.get("role") == "assistant":
                    print(f"\n[AI]: {msg['content']}")
            last_displayed_message_idx = len(messages) - 1

        # 2. Check for Interrupts
        if state.tasks and state.tasks[0].interrupts:
            interrupt_info = state.tasks[0].interrupts[0].value

            # --- FIX: Safely handle both String and Dict interrupts ---
            if isinstance(interrupt_info, dict):
                ai_message = interrupt_info.get('content', 'No message content')
                options = interrupt_info.get('available_options', [])
            else:
                ai_message = interrupt_info
                options = []

            print(f"\n[AI]: {ai_message}")
            if options:
                print(f"Options: {', '.join(options)}")
            # -------------------------------------------------------

            user_input = input("\n[YOU]: ")
            print(f"[YOU]: {user_input}")

            # Resume the graph with the user's input
            graph.invoke(Command(resume=user_input), config=config)

        # 3. Check if the graph has finished
        elif not state.next:
            # Before ending, check if there's a final assistant message to print
            if state.values and state.values.get('messages') and state.values['messages'][-1]["role"] == "assistant":
                if last_displayed_message_idx < len(state.values['messages']) - 1:
                    print(f"\n[AI]: {state.values['messages'][-1]['content']}")
            print("\n--- ⚑⚑ Session Ended ---")
            break

        # 4. If nodes are pending but no interrupt, let them run (the gas pedal)
        else:
            graph.invoke(None, config=config)

# IMPORTANT: Set reset=True only when you want to wipe the history.
# Set it to False to actually continue the conversation!
run_booking_session(booking_graph, reset=True)

Conclusion:

For the Conclusion stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.

References:

For the References stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments. Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness. For the References stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

Operational checklist

For the Operational checklist stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state.

Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph.

Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.

Write a short runbook: how to rotate keys, how to drain the queue, how to roll back the last ingest.

Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.

Before promoting the stack, freeze versions, capture a golden transcript for the critical path, and confirm rollback steps. Shared environments need rate limits, tenancy checks, and a clear owner for secret rotation. Prefer boring reliability over clever one-off demos.

Batch note for 9744ef4a5b25: keep provider keys out of the repo, set a per-session token ceiling, and store transcripts next to the eval fixtures so later model swaps stay comparable.