Museum Booking Chatbot

🎬 Streamlit App Screen Recording

📱 Chatbot Landing Interface

💡 The Problem

Tourists and locals wanting to visit Indian museums face fragmented information — scattered across multiple websites with no unified way to search, compare, and book tickets. Museums need 24/7 booking capability but can't afford dedicated staff for every inquiry.

🎯 The Solution

A conversational AI agent that understands natural language queries about museums. Users can search by city or theme, get detailed info (timings, prices, facilities), and complete ticket bookings — all within a single chat interface. The agent remembers context across turns, enabling natural multi-step workflows like search → details → book.

⚡ Key Features

🔍

Fuzzy-Matched Search

Typo-tolerant museum search using Python's difflib. Users can type "Kalkutta" and still find museums in Kolkata.

🧠

Stateful LangGraph Agent

ReAct agent with MemorySaver checkpointing — remembers the entire conversation context for multi-turn workflows.

🎟️

End-to-End Booking

Search museums → get details → book tickets with automatic price calculation for adults and children.

✂️

Context History Trimming

Intelligent message trimming keeps conversations within the LLM's 8192 token context window without losing critical context.

🆕

New Chat / Thread Support

Users can start fresh conversations with a new thread ID, keeping previous booking history intact while starting a clean chat.

♿

Wheelchair Accessibility Filter

Dedicated search parameter for wheelchair-accessible museums, making the tool inclusive for all users.

🏗️ Architecture

👤 User Message Streamlit Chat UI

→

🧠 LangGraph ReAct Agent qwen2.5(base) via Ollama

→

🔧 Tool Selection Search / Details / Book

→

🗄️ MySQL Database Connection Pool

Agent & Tools (core/)

llm.py — LangGraph ReAct agent with history trimming & checkpointing
tools.py — Three LangChain tools: search_museums, get_museum_details, book_ticket
db.py — MySQL connection pool for handling concurrent requests

Frontend & Entry Points

app.py — Streamlit web UI with chat interface & new conversation button
terminal.py — CLI entry point for terminal-based interaction

💬 Interactive Conversation Walkthroughs

Step-by-step walkthroughs of the chatbot performing museum discovery, ticket booking, and error handling.

🔄 Flow 1: Search & Successful Booking (Pune)

Step 1: Discovering Museums

The user searches for "list of museums in pune", prompting the chatbot to retrieve matching entries from the MySQL database.
Step 2: Exploring Museum Details

The chatbot provides comprehensive information including address, timings, entry fees, and facilities for the selected museum.
Step 3: Booking Request

The user asks to book tickets. The chatbot automatically extracts parameters and queries for confirmation with a pricing summary.
Step 4: Ticket Confirmed

Upon user confirmation, the system creates a secure booking entry in the database and returns a validation receipt.

🔄 Flow 2: Multi-Turn Corrections & Handling Edge Cases

Step 1: Regional Search

Searching for "museums in goa" lists all matching options. The system displays 10 cultural sites in the region.
Step 2: Cross-Region Clarification

When requesting a museum outside Goa (Heritage Transport Museum), the bot identifies the correct state (Haryana) and continues.
Step 3: Name & State Disambiguation

For ambiguous names like "Darpan Art Gallery", the agent checks matching database entries and asks the user to confirm the state.
Step 4: Date Validation Error

If the booking date is set in the past, the database transaction aborts and the agent notifies the user of the validation failure.

🧩 Agent Tools

`search_museums_tool`

Search museums by city name, theme, or keyword. Supports wheelchair accessibility filtering. Uses fuzzy matching to handle typos and partial names.

query needs_wheelchair

`get_museum_details_tool`

Retrieve complete information for a specific museum — address, timings, prices, facilities, overview, and closed days.

museum_name city

`book_ticket_tool`

Book tickets for a specific museum, date, and party size. Automatically calculates total price based on adult and child counts.

museum_name city booking_date adult_count child_count

🛠️ Tech Stack

LLM

qwen2.5 local via Ollama)

Agent Framework

LangGraph (ReAct + checkpointing)

Tool Bindings

LangChain

Frontend

Streamlit

Database

MySQL + Connection Pool

Fuzzy Matching

Python difflib

🧩 Challenges & Learnings

Preventing LLM Tool-Call Hallucinations

The qwen2.5 model would sometimes fabricate museum data instead of calling the database tools. Solved by refining the system prompt with strict instructions to always use tools for museum queries, and building an evaluation dataset to test accuracy across different query types.

Context Window Overflow in Long Conversations

Booking workflows can span 10+ messages, easily overflowing the 8192 token limit. Implemented trim_messages with start_on="human" to intelligently prune older messages while keeping the conversation coherent. Made the max history configurable via MAX_HISTORY_MESSAGES.

Handling User Typos in Museum Names

Users type museum names in many ways — "indian museum", "Indian Musuem", "indain museum". Implemented fuzzy matching with Python's difflib to find the closest match and return relevant results even with significant spelling errors.