MealFeel

A personal wellness app that helps you log meals quickly and explore connections between what you eat and how you feel.

• Meal Tracking: Log meals in natural language or via photo. View and edit your history.
• Insight Chatbot: Ask questions like "What did I eat last week?" or "Which foods might be linked to my bloating?"

---

What does MealFeel do?

Meal Tracking

• Log meals fast: Type a natural language sentence (e.g., "For lunch today I had a hamburger, french fries and a wedge salad and afterwards felt bloated"), or upload a photo to auto-extract a description and symptoms.
• Automatic structure: The app parses your text into date, meal type, dishes, and optional symptoms.
• Ingredient awareness: It expands dish names into likely ingredients (e.g., hamburger → beef, tomato, onion, pickles, potatoes, vegetable oil, iceburg lettuce).
• History and edits: Browse your logged meals, open a meal, and edit details later.

Insight Chatbot

• Meal recall: "What did I eat yesterday?", "When did I last have eggs?"
• Symptom patterns: "What foods make me feel bloated?", "Which ingredients show up when I have headaches?"
• Safe guidance: Results emphasize patterns and correlations, not medical advice.

---

How it works (technical overview)

Architecture

• Client (Vite + React + MUI): UI components for logging meals, viewing history, and chatting.
• Server (Node + Express + PostgreSQL): REST API for meals and a chat endpoint that runs RAG-style analysis.
• AI: OpenAI (GPT-4o) for structured parsing, ingredient inference, and safe response generation; OpenAI embeddings for semantic search.
• Observability: Langfuse tracing around AI calls.

Key flows

• Meal logging (POST /api/meals):

  1. Parse free‑text into structured fields via Instructor + GPT-4o (schema‑validated with Zod)
  2. Infer likely ingredients from dishes using GPT-4o
  3. Generate a clinically toned meal_text summary
  4. Create an embedding for semantic recall and store everything in Postgres

• Meal recall (POST /api/chat with intent meal_recall):

  • Parses time expressions (e.g., "last week"), fetches relevant meals, uses embeddings to rank by similarity, and formats a readable summary preserving line breaks.

• Symptom analysis (POST /api/chat with intent symptom_analysis):

  • Finds meals mentioning symptoms, runs ingredient frequency/intersection analysis, and presents possible correlations with disclaimers.

---

Setup

Prerequisites

• Node.js 18+
• PostgreSQL 14+
• OpenAI API key

Environment variables (server)

Create a .env file in server/:

PORT=3001
OPENAI_API_KEY=your_openai_key
DB_HOST=localhost
DB_PORT=5432
DB_USER=postgres
DB_PASSWORD=postgres
DB_NAME=wellness_tracker
LANGFUSE_PUBLIC_KEY=optional
LANGFUSE_SECRET_KEY=optional
LANGFUSE_HOST=optional

Install & run

In two terminals:

# Terminal 1 – server
cd server
npm install
npm run dev

# Terminal 2 – client
cd client
npm install
npm run dev

Client dev server proxies '/api' to http://localhost:3001.

Database

• Import the dump in db/wellness_tracker_dump.sql to create tables and seed sample data.
• The meals table stores dishes and ingredients as arrays and a meal_text string. An embedding vector is stored as text in this POC and used for similarity with PostgreSQL operators in queries.
• Proof of Concept project uses data from a single user. Data taken from kaggle and transformed using google colab

---

API

Meals

• GET /api/meals → list meals (newest first)
• GET /api/meals/:id → fetch one meal
• POST /api/meals → create a meal
  • Body: { "text": string, "session_id?": string }
  • Response: full meal record plus debug session info
• PATCH /api/meals/:id → update a meal

Chatbot

• POST /api/chat
  • Body: { "query": string, "user_id?": number, "session_id?": string }
  • Intents: meal_recall, symptom_analysis, general
  • Returns: intent classification, number of meals considered, and a formatted answer

---

Frontend (selected components)

• MealTracker.tsx: natural‑language input, photo upload modal, and a list of logged meals.
• InsightFinder.tsx: chat interface to send queries and display AI responses (markdown‑rendered).

---

Safety & disclaimers

• This app does not provide medical advice.
• Pattern findings indicate correlation, not causation.
• Consult a healthcare professional for medical concerns.