BaristaBot: AI-Powered Cafe Ordering System

A conversational AI ordering system built with LangGraph and Google Gemini that implements semantic menu search, order state management, and automated response quality control.

Overview

This project implements a multi-agent conversational system for cafe ordering. The system uses vector-based semantic search for menu retrieval, maintains stateful order information throughout the conversation, and evaluates AI responses before presenting them to users.

Technical Implementation

Core Architecture

• LangGraph: State graph implementation for conversation flow control
• Google Gemini (gemini-1.5-flash): Language model for natural language understanding and generation
• FAISS: Vector database for semantic menu search with cosine similarity
• LangChain: Framework integration and tooling

Key Components

Vector-Based Menu Search

• FAISS vector store with GoogleGenerativeAI embeddings
• Semantic search supporting natural language queries
• k=4 nearest neighbor retrieval for relevant menu items

Order State Management

• Persistent order state across conversation turns
• Operations: add_to_order, confirm_order, get_order, clear_order, place_order
• Explicit confirmation required before finalizing orders

Response Quality Control

• Automated evaluation of AI responses before delivery
• Correction system for responses not meeting quality standards
• Verification against menu constraints and clarity requirements

Conversation Flow

• Five-node graph: chatbot, human, tools, ordering, evaluator
• Dynamic routing based on tool calls and message content
• Recursion limit of 100 for extended conversations

Attribution

Based on the 5-Day Gen AI Intensive Course original code from Day 3 - Building an Agent with LangGraph.

Course Developers: Addison Howard, Brenda Flynn, Myles O'Neill, Nate, Polong Lin

Competition: Gen AI Intensive Course Capstone 2025Q1

Enhancements Over Original

1. Vector Database Integration: Replaced hardcoded menu with FAISS vector store for scalable, semantic menu search
2. RAG Implementation: Dynamic retrieval of menu information based on query context
3. Quality Evaluation System: LLM-based response validation and correction before user presentation
4. Enhanced Tool Organization: Separation of informational tools (menu search) from state-modifying tools (order management)

Dependencies

langgraph==0.3.21
langchain-google-genai==2.1.2
langgraph-prebuilt==0.1.7
langchain_community
faiss-cpu

Configuration

Requires GOOGLE_API_KEY environment variable for Gemini API access. In Kaggle environments, this should be stored as a Kaggle secret.

Usage

The system initiates with a welcome message and enters a conversation loop. Users can:

• Search menu items using natural language
• Add items to order with specifications (milk type, sweetness, temperature)
• View current order
• Confirm and place orders
• Exit conversation with 'q' command

Technical Notes

Vector Search Implementation

• Embedding model: models/embedding-001
• Distance metric: Cosine similarity
• Retrieval configuration: Top 4 results

State Management

• TypedDict-based state with annotated message history
• Order state persists across graph invocations
• Tool call metadata preserved during evaluation corrections

Graph Compilation

• Conditional edges for dynamic routing
• Tool execution loops back to chatbot node
• Evaluator positioned between chatbot and routing logic

License

Apache License 2.0