Skip to content

AI-Powered Meal Planning Assistant

UI Streamlit

Portfolio Best Practices

This is a portfolio project demonstrating RAG implementation and full-stack AI development. Key highlights:

  • End-to-end RAG pipeline with semantic search
  • PostgreSQL + pgvector for production-grade vector storage
  • Structured outputs using Pydantic models
  • Docker-based deployment with service orchestration
  • Clean separation of concerns (UI, business logic, data layer)
  • Scalable architecture supporting 2M+ recipe embeddings

Project Summary

Project: MealPrep - AI Meal Planning Assistant
Repository: GitHub
Category: Personal AI Application

Key Features:

  • Semantic search across 2.2M+ recipes using RAG
  • Single meal suggestions based on available ingredients
  • Multi-day meal plan generation with shopping lists
  • Real-time meal regeneration with diversity enforcement
  • PostgreSQL database with meal plan persistence
  • Built from scratch in 1 week

An intelligent meal planning application that combines Retrieval-Augmented Generation (RAG) with Large Language Models to deliver personalized meal suggestions and automated weekly meal plans based on available ingredients and dietary preferences.

Challenge

Planning meals daily is time-consuming and often results in repetitive choices. Existing meal planning apps lack:

  • Intelligent suggestions based on what you already have
  • True variety - they suggest similar meals repeatedly
  • Flexibility - rigid plans that don't adapt to rejections
  • Semantic understanding - can't find similar recipes or adapt to preferences

The goal was to build an AI-powered solution that understands context, ensures variety, and learns from user preferences while leveraging a massive recipe database.

Our Approach

Developed a full-stack RAG application with two distinct modes:

Mode 1: Single Meal Suggestion

  • Ingredient-based semantic search through 2M+ recipes
  • RAG pipeline retrieves similar recipes as context
  • LLM (OpenAI GPT-5-mini) generates personalized suggestions considering dietary preferences
  • Accept/reject workflow with automatic diversity enforcement

Mode 2: Multi-Day Meal Planning

  • Automated generation of X day meal plans
  • Individual meal acceptance/rejection with on-the-fly regeneration
  • Shopping list aggregation from accepted meals
  • Plan persistence with database storage

Technical Implementation:

  • Embedded 2.2M recipes using sentence-transformers (text-embedding-3-small)
  • Vector similarity search with pgvector and PostgreSQL
  • Structured outputs using Pydantic for reliable parsing
  • Prompt engineering for diversity and creativity

Results & Impact

  • Successfully processes 2.2M+ recipe embeddings
  • Sub-second semantic search response times
  • Reliable structured outputs using Pydantic validation
  • Meal plan generation in under 30 seconds for 7-day plans
  • Docker deployment with persistent storage
  • Clean architecture enabling easy feature additions

Solution Overview

Architecture

┌─────────────┐
│  Streamlit  │  UI Layer
│     UI      │
└──────┬──────┘
       │
┌──────▼──────────┐
│  Meal Service   │  Business Logic
│   (RAG Flow)    │
└──────┬──────────┘
       │
┌──────▼──────────┐     ┌─────────────┐
│ Vector Store    │────▶│  PostgreSQL │
│  (pgvector)     │     │  + pgvector │
└─────────────────┘     └─────────────┘
       │
┌──────▼──────────┐
│   Claude API    │  LLM Generation
│   (Sonnet 4)    │
└─────────────────┘

RAG Pipeline

  1. Indexing Phase:

    • Load 2.2M recipes from Kaggle dataset
    • Generate embeddings using sentence-transformers
    • Store in PostgreSQL with pgvector extension

  2. Retrieval Phase:

    • User inputs ingredients
    • Semantic search finds similar recipes (cosine similarity)
    • Top-k results retrieved as context

  3. Generation Phase:

    • Similar recipes passed to OpenAI as system context
    • Prompt includes ingredients, dietary preferences, excluded meals
    • Structured output ensures consistent parsing

  4. Diversity Enforcement:

    • Recent meals tracked in database
    • Rejected meals added to exclusion list
    • Prompt explicitly instructs variety across cuisines and methods

Tech Stack

AI & Machine Learning

  • OpenAI GPT-5-mini for meal generation
  • sentence-transformers (text-embedding-3-small) for embeddings
  • pgvector for vector similarity search
  • Custom RAG pipeline implementation

Backend

  • Python 3.12
  • Streamlit for rapid UI development
  • PostgreSQL with pgvector extension
  • Pydantic for data validation
  • psycopg2 for database connectivity

Data Processing

  • Pandas for recipe data manipulation (added diet classifications)
  • 2.2M recipe dataset from Kaggle

Infrastructure

  • Docker & Docker Compose for service orchestration
  • uv for fast Python package management
  • python-dotenv for configuration management

Development

  • Git for version control
  • Structured project layout with separation of concerns

Additional Context

  • Timeline: 1 week
  • Team Size: Solo project
  • Role: Full-stack AI Engineer
  • Dataset: 2.2M+ recipes from Kaggle
  • Learning Focus: Production RAG implementation, vector databases, structured outputs

Future Enhancements:

  • Shopping list integration with grocery apps
  • MCP (Model Context Protocol) integration with growing RAG database (see below)
  • Advanced text chunking for recipe processing
  • PDF cookbook parsing
  • Nutritional tracking
  • Calendar integration for meal scheduling

Learnings & Insights

Technical Challenges Solved:

  1. Scale: Efficiently embedding and searching 2M+ recipes
  2. Parsing Reliability: Structured outputs improved parsing
  3. Diversity: Prompt engineering crucial for non-repetitive suggestions
  4. State Management: Streamlit session state for complex workflows

Key Takeaways:

  • RAG quality depends heavily on retrieval strategy
  • Prompt engineering is as important as model choice
  • Structured outputs essential for production reliability
  • Docker simplifies local development and deployment

  • Let's have a virtual coffee together!

    Want to see if we're a match? Let's have a chat and find out. Schedule a free 30-minute strategy session to discuss your AI challenges and explore how we can work together.

    Book Free Intro Call