App Overview

Medley is a Hims-branded hair loss consultation chat demo built with the Foundation Models framework and SwiftUI to provide an on-device, interactive user experience.

Medley AI Demo - Watch Video

Project Memo

Core Architecture

Data Schema (DataSchema.swift)

• Customizable JSON question structure to easily control the AI consultation flow.
• Each question contains: prompt, question type (single/multiple choice, free text), predefined response buttons, additional info, and next-question rules
• Schema is loaded and indexed in the Foundation Model session.

Conversation Model (ConversationModel.swift)

• Protocol-based design; currently powered by Foundation Models (Apple's on-device LLM)
• Other implementations (ChatGPT, Claude, etc.) that conform to the protocol can be swapped in
• Streams text responses, handles answer mapping, and determines flow progression
• Generates context-aware responses using the schema and conversation history
• Uses the LLM to map the user's open-ended response with a predefined option from the schema

Chat ViewModel (ChatViewModel.swift)

• Orchestrates conversation flow: loads schema, manages messages array, tracks currentQuestion, and holds collected data
• For each turn in the conversation: streams the model's response, maps user input to structured data, advances to next question
• Handles three types of streamed messages: opening, acknowledgment + next question, or info summary + question
• Answer Mapping: auto-detects exact matches or uses LLM to categorize open-ended responses into schema options

UI Flow

Chat Screen

• Title + Reset button (clears chat and resets data)
• Scrollable message thread (auto-scrolls to newest message)
• Predefined response chips (contextual buttons from current question)
• Text input bar for open-ended responses from the user

Message Rendering (ChatView.swift)

• User messages are right-aligned with a dark bubble
• Model messages are left-aligned with a light bubble with streaming animation
• Messages auto-scroll to bottom as new content arrives

Data Collection

• Each user response is validated/mapped to the schema's structured data (StructuredConsult.swift)
• Multiple-choice questions append to arrays; single-choice replace values

Completion Flow

• When final question is answered, isComplete flag triggers a "Next" button
• Tapping "Next" shows ResultsView with the collected data as pretty-printed JSON

Key Design Patterns

• Observation: ViewModel uses @Observable for reactive updates
• Fallbacks: If LLM call fails, app falls back to schema text gracefully
• User answers are simple string IDs but the model can return more complex structured data using the Generable protocol