🍽️ LA Luxury Restaurant Recommendation System

An End-to-End NLP & Semantic Search Pipeline for Upscale Dining in Los Angeles

---

📌 Project Overview

This project is a production-style, multi-stage AI pipeline that transforms a curated dataset of 71 upscale and Michelin-recognized Los Angeles restaurants into an intelligent, conversational recommendation engine. Users can query the system in plain English — "romantic rooftop Italian restaurant in Beverly Hills" — and receive semantically ranked, filter-refined recommendations powered by embeddings, zero-shot classification, emotion analysis, and a large language model.

Built as an adaptation of the FreeCodeCamp Semantic Book Recommender tutorial, this project translates that framework into the restaurant domain, extending it with multi-notebook architecture, metadata enrichment, multi-label NLP classification, and an interactive Gradio front end.

Scope: Los Angeles County — including Downtown LA, Beverly Hills, West Hollywood, Santa Monica, Hollywood, Culver City, Malibu, and surrounding neighborhoods. Restaurants span Michelin 3-Star down to Michelin Selected and high-end non-Michelin establishments.

---

🎬 Project Demo

VS Code Project Structure & Gradio Dashboard

Project Structure & Dashboard Code

Gradio Dashboard UI

---

🧠 What This Project Demonstrates

This is not a tutorial copy — it is a domain-adapted, multi-notebook AI engineering project that showcases:

Skill Area	What Was Applied
Data Engineering	CSV ingestion, multi-column inspection, whitespace normalization, address correction, metadata column generation
Semantic Search	Sentence transformer embeddings + ChromaDB vector store with persisted storage
NLP Classification	Zero-shot classification via facebook/bart-large-mnli for cuisine, occasion, and vibe
Emotion & Sentiment Analysis	Multi-label emotion scoring via distilbert-base-uncased-emotion + dominant emotion extraction
LLM Integration	Claude Sonnet via Anthropic API for natural language recommendation synthesis
Pipeline Architecture	4-notebook modular design with shared CSV artifacts between stages
UI Development	Gradio dashboard with real-time query, multi-filter support, and restaurant cards
Software Engineering	.env secret management, .gitignore, YAML-based version control & script validation

---

🗂️ Project Structure

RestaurantRecommenderLLM/
│
├── data/
│   ├── cleaned_restaurant_List.csv              # Original cleaned dataset (71 restaurants)
│   ├── cleaned_restaurants_final.csv            # + restaurant_metadata column
│   ├── restaurants_with_classifications.csv     # + cuisine group, dining format, occasion, vibe
│   ├── restaurants_with_emotions.csv            # + 7 emotion scores, sentiment, dining mood
│   └── tagged_restaurant_descriptions.txt       # Formatted text for ChromaDB ingestion
│
├── jupyter_scripts/
│   ├── 1_Restaurant_Data_Cleanup.ipynb          # Stage 1: Data inspection & enrichment
│   ├── 2_Restaurant_Vector_Search.ipynb         # Stage 2: Embeddings & semantic retrieval
│   ├── 3_Restaurant_Text_Classification.ipynb   # Stage 3: Zero-shot NLP classification
│   └── 4_Restaurant_Sentiment_Analysis.ipynb    # Stage 4: Emotion & sentiment scoring
│
├── gradio_app/
│   └── GradioDashboard.py                       # Interactive web UI
│
├── chroma_restaurants/                          # Persisted ChromaDB vector store
│
├── photos/
│   ├── SamplePic1.png
│   └── SamplePic2.png
│
├── .env                                         # API keys (never committed)
├── .gitignore
└── README.md

---

⚙️ Multi-Notebook Pipeline Architecture

Each Jupyter notebook is a self-contained stage that reads from and writes to shared CSV artifacts. This modular design allows each component to be tested, updated, or swapped independently — a hallmark of production-grade ML pipelines.

Raw CSV
   │
   ▼
[Notebook 1] Data Cleanup & Metadata Generation
   → cleaned_restaurants_final.csv  (+restaurant_metadata)
   │
   ▼
[Notebook 2] Vector Embeddings & Semantic Search
   → chroma_restaurants/ (persisted ChromaDB)
   │
   ▼
[Notebook 3] Zero-Shot Text Classification
   → restaurants_with_classifications.csv  (+cuisine_group, dining_format, occasion, vibe)
   │
   ▼
[Notebook 4] Emotion & Sentiment Analysis
   → restaurants_with_emotions.csv  (+7 emotion scores, dominant_emotion, dining_mood)
   │
   ▼
[Gradio Dashboard] — Combines all stages for live querying

---

📓 Notebook Breakdown

Notebook 1 — Data Cleanup & Metadata Enrichment

• Inspected all 14 columns across 71 restaurant records
• Resolved trailing whitespace in Address and Description columns
• Corrected a missing comma in Mastro's Ocean Club address
• Fixed an incorrect zip code for Morihiro (Echo Park location)
• Confirmed intentional multi-location entries for Pine & Crane and Badmaash
• Generated a restaurant_metadata column combining structured fields into enriched descriptive text for downstream NLP tasks

Notebook 2 — Vector Search with ChromaDB

• Loaded tagged_restaurant_descriptions.txt into ChromaDB using all-MiniLM-L6-v2 sentence embeddings (HuggingFace, local inference — no API cost)
• Implemented a candidate pool strategy to prevent filter depletion when combining semantic search with metadata filters
• Core retrieval function accepts 6 parameters: query, top_k, michelin_filter, price_filter, atmosphere_filter, rooftop_only
• Includes 10 diverse test queries and a scored similarity debugging cell
• ChromaDB store persisted to ./chroma_restaurants/ for reuse across sessions

Notebook 3 — Zero-Shot Text Classification

• Used facebook/bart-large-mnli to classify each restaurant across three dimensions without any labeled training data:
  • Cuisine Group — e.g., Japanese, Italian, Seafood, Mexican
  • Dining Format — e.g., Tasting Menu / Omakase, Full Service, Counter Dining
  • Best For (Occasion) — e.g., Special Occasion, Date Night, Business Dinner
  • Vibe — e.g., Refined & Elegant, Hip & Trendy, Casual & Relaxed
• Confidence scores stored alongside each label for downstream filtering
• Output: restaurants_with_classifications.csv (21 columns)

Notebook 4 — Emotion & Sentiment Analysis

• Applied bhadresh-savani/distilbert-base-uncased-emotion to restaurant_metadata field
• Extracted 7 per-restaurant emotion probability scores: anger, disgust, fear, joy, neutral, sadness, surprise
• Derived dominant_emotion and dining_mood labels from score distributions
• Added overall_sentiment (Positive / Neutral / Negative) using a secondary sentiment classifier
• Output: restaurants_with_emotions.csv (32 columns) — the final enriched dataset powering the Gradio dashboard

---

🤖 Models & Tools Used

Embeddings

Model	Purpose
sentence-transformers/all-MiniLM-L6-v2	Dense vector embeddings for semantic similarity search

Classification

Model	Purpose
facebook/bart-large-mnli	Zero-shot multi-label classification (cuisine, format, occasion, vibe)

Sentiment & Emotion

Model	Purpose
bhadresh-savani/distilbert-base-uncased-emotion	7-class emotion scoring from text
Secondary sentiment classifier	Positive / Neutral / Negative overall sentiment

LLM

Model	Purpose
Claude Sonnet (Anthropic API)	Natural language recommendation synthesis and query understanding

Infrastructure

Tool	Role
ChromaDB	Local persisted vector store
Gradio	Web dashboard UI
pandas	Data manipulation across all pipeline stages
python-dotenv	Secure API key management via .env
YAML	Version control, script validation & pipeline verification

---

🗄️ Dataset

• 71 curated restaurants across Los Angeles County
• Covers Michelin 3-Star, 2-Star, 1-Star, Bib Gourmand, Michelin Selected, and high-end non-Michelin establishments
• Data sourced from: Michelin Guide, OpenTable, Time Out LA, restaurant websites
• Each restaurant record contains 14 base fields, enriched to 32 columns by end of pipeline

Sample enriched columns in final dataset:

Column	Description
restaurant_metadata	Enriched descriptive text for NLP tasks
simple_cuisine_group	Classified cuisine category
dining_format	Tasting Menu, Full Service, Counter, etc.
predicted_occasion	Best For — Special Occasion, Date Night, etc.
predicted_vibe	Refined & Elegant, Hip & Trendy, etc.
emotion_joy / emotion_neutral / ...	Per-restaurant emotion probability scores
dominant_emotion	Highest-scoring emotion label
dining_mood	Human-readable mood derived from emotion profile
overall_sentiment	Positive / Neutral / Negative

---

🔄 Version Control & Validation

This project uses a YAML configuration file to manage script versions, track pipeline stage dependencies, and validate that each notebook's output schema matches expected column definitions before the next stage ingests it. This ensures pipeline integrity when notebooks are updated or re-run out of order — a critical safeguard in multi-stage ML workflows.

---

🚀 Getting Started

Prerequisites

• Python 3.11+
• pip / virtualenv
• Anthropic API key

Installation

git clone https://github.com/papasmurf79/RestaurantRecommenderLLM.git
cd RestaurantRecommenderLLM
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

Environment Setup

Create a .env file in the project root:

ANTHROPIC_API_KEY=your_anthropic_api_key_here

Running the Pipeline

Run notebooks in order from the jupyter_scripts/ directory:

1_Restaurant_Data_Cleanup.ipynb
2_Restaurant_Vector_Search.ipynb
3_Restaurant_Text_Classification.ipynb
4_Restaurant_Sentiment_Analysis.ipynb

Launching the Dashboard

cd gradio_app
python GradioDashboard.py

---

🔮 Planned Enhancements

• Natural language filter parsing using Claude (e.g., "under $200 per person" → price_filter)
• Map integration for neighborhood-based browsing
• OpenTable / Tock reservation link embedding
• Voice AI agent to read back restaurant information
• User preference memory across sessions
• Expansion to Orange County and San Diego restaurant datasets
• Streamlit or React frontend migration
• CI/CD pipeline for automated data refresh

---

📚 References & Inspiration

• FreeCodeCamp — Build a Semantic Book Recommender Using an LLM and Python
• Supercharge Restaurant Recommendations with LLM and Qdrant — Medium
• Restaurant Recommender LLM — DoraHacks
• Anthropic Claude API Documentation
• ChromaDB Documentation
• HuggingFace Transformers
• Michelin Guide — Los Angeles

---

👤 Author

Papasmurf AI Engineer | NLP & Semantic Search | Los Angeles, CA

---

Built with curiosity, sashimi, and a deep respect for the Michelin Guide.