Tapestry - Personal Knowledge Graph & Thought Recorder

A private, intelligent iOS app for recording and exploring your thoughts through voice with AI-powered insights

Key Features

Voice-First Thought Capture

• One-tap audio recording with real-time waveform visualization
• On-device speech recognition using Apple's Speech framework
• Automatic transcription with confidence scoring
• Haptic feedback and intuitive UI design

AI-Powered Oracle Assistant

• Local LLM powered by Leap iOS SDK with LFM2-1.2B model
• Contextual memory exploration and pattern analysis
• Personal chat interface that understands your thought patterns
• No cloud dependency - all AI processing happens on-device

Advanced Knowledge Graph Database

• Kuzu Graph Database integration for complex relationship modeling
• MCP (Model Context Protocol) Server for structured thought queries
• Semantic search across thoughts, emotions, people, and activities
• Graph-based pattern detection and insight generation

Intelligent Search & Analysis

• Natural language queries: "What makes me stressed?", "Fun memories with friends"
• Emotion tracking and sentiment analysis over time
• Health correlation analysis (HealthKit integration)
• Location-based thought clustering
• Keyword networks and topic modeling

Rich Analytics & Visualization

• Interactive 3D thought network using Three.js
• Temporal emotion timeline with intensity tracking
• Swift Charts for statistical insights
• Pattern analysis across time, location, and context

Privacy & Security First

• 100% Local Processing - no cloud sync or external APIs
• Face ID/Touch ID app lock protection
• File-level encryption (NSFileProtectionComplete)
• On-device speech recognition and AI inference
• Your thoughts never leave your device

LFM2-1.2B: The AI Engine That Powers Everything

At the heart of Tapestry lies LFM2-1.2B, a remarkably versatile 1.2 billion parameter language model that demonstrates how effective privacy-first AI can be. Despite its compact size, LFM2 drives multiple sophisticated AI capabilities throughout the app:

Constrained Generation for Structured Extraction

• Keyword Extraction: Precisely identifies key concepts from voice transcriptions using constrained generation to ensure consistent, structured output
• Entity Recognition: Extracts people, places, emotions, and activities with high accuracy while maintaining strict output formats
• Metadata Generation: Creates structured thought metadata including sentiment scores, confidence ratings, and contextual tags

Conversational Oracle Agent

• Memory Synthesis: Acts as your personal memory companion, weaving together insights from your thought patterns with natural, empathetic responses
• Context Awareness: Maintains conversation context while drawing from your personal knowledge graph to provide meaningful, personalized insights
• Adaptive Personality: Develops understanding of your communication style and emotional patterns over time

Function Calling & MCP Integration

• Tool Selection: Intelligently chooses appropriate MCP tools based on natural language queries ("What makes me stressed?" → emotion pattern analysis tools)
• Parameter Generation: Converts conversational requests into precise database query parameters
• Result Interpretation: Transforms raw database results into meaningful, conversational insights

Cypher Query Generation

• Graph Database Queries: Generates complex Cypher queries for the Kuzu knowledge graph, enabling sophisticated relationship exploration
• Dynamic Query Optimization: Adapts query complexity based on data patterns and user intent
• Multi-hop Reasoning: Creates compound queries that traverse multiple relationship types to uncover hidden insights

Pattern Analysis & Insights

• Temporal Analysis: Identifies trends in emotions, activities, and thoughts across different time periods
• Correlation Discovery: Detects relationships between health metrics, locations, people, and emotional states
• Narrative Generation: Transforms statistical patterns into compelling, personal insights

Why LFM2-1.2B is Perfect for Personal AI

The genius of using LFM2-1.2B lies in its optimal balance between capability and privacy. While larger models require cloud processing or massive local resources, LFM2's compact architecture enables:

• Real-time Performance: Sub-second response times for all AI operations
• Battery Efficiency: Minimal impact on device battery life during extended use
• Device Compatibility: Runs smoothly on iPhone models back to the A12 Bionic chip
• Complete Privacy: Zero data leaves your device - every AI operation happens locally
• Multi-modal Intelligence: Single model handles text generation, structured extraction, function calling, and conversational AI

This makes Tapestry uniquely powerful among personal AI apps - you get enterprise-grade intelligence with consumer-friendly privacy and performance. LFM2 proves that you don't need massive models or cloud dependencies to create truly intelligent, personalized experiences.

Technical Architecture

Core Stack

• UI: SwiftUI with iOS 17+ features
• Audio: AVFoundation (m4a/AAC, optimized for voice)
• AI/LLM: Leap iOS SDK with LFM2-1.2B local model inference
• Graph Database: Kuzu (embedded graph database)
• Storage: Core Data + Kuzu dual storage system
• Search: Vector embeddings with NaturalLanguage framework
• 3D Visualization: WKWebView + Three.js + 3D Force Graph

Advanced Features

• MCP Server: Model Context Protocol for structured AI interactions
• Semantic Embeddings: Vector similarity for thought relationships
• Health Integration: HealthKit correlation analysis
• Location Context: CoreLocation for geographical insights
• Biometric Security: LocalAuthentication framework integration

Project Structure

Tapestry/
├── TapestryApp.swift                   # App entry point with MCP server init
├── Models/
│   ├── Tapestry.xcdatamodeld          # Core Data schema
│   ├── PersistenceController.swift   # Core Data + vector search
│   ├── ChatMessage.swift             # Chat interface models
│   └── ThoughtPatternModels.swift    # Graph analysis models
├── Views/
│   ├── ContentView.swift             # Main app navigation
│   ├── RecordView.swift              # Voice recording interface  
│   ├── ChatView.swift                # Oracle AI chat interface
│   ├── TimelineView.swift            # Thought timeline browser
│   ├── ThoughtPatternExplorationView.swift # Graph visualization
│   └── AnalyzeView.swift             # Analytics dashboard
├── Services/
│   ├── AudioManager.swift            # Audio recording & playback
│   ├── SpeechManager.swift           # Speech recognition + embeddings
│   ├── ChatStore.swift               # AI conversation management
│   ├── KuzuManager.swift             # Graph database operations
│   ├── MCPThoughtServer.swift        # MCP protocol server
│   ├── HealthKitManager.swift        # Health data integration
│   ├── LocationManager.swift         # Location context collection
│   └── BiometricAuthManager.swift    # Security & authentication
├── MCP Tools/                         # Specialized database query tools
│   ├── MCPThoughtServerTools.swift       # Core search & retrieval
│   ├── MCPThoughtServerAnalysisTools.swift # Pattern analysis
│   ├── MCPThoughtServerContextTools.swift  # Location & health correlation
│   └── MCPThoughtServerStatTools.swift     # Statistical insights
├── Graph/
│   └── index.html                    # 3D force graph visualization
└── Resources/
    ├── LFM2-1.2B-8da4w_output_8da8w-seq_4096.bundle # Local LLM model
    └── Assets.xcassets               # App icons & images

Setup Instructions

Prerequisites

• Xcode 15.0+
• iOS 17.0+ target device
• ~2GB free space (for LFM2-1.2B model and app data)

Installation

1. Clone & Open Project

   git clone <repository-url>
   cd ThoughtDiary
   open Tapestry.xcodeproj

2. Install Dependencies (Auto-resolved via SPM)

   • Kuzu Swift: https://github.com/kuzudb/kuzu-swift.git
   • Leap iOS SDK: https://github.com/liquid4all/leap-ios

3. Add 3D Graph Assets to Graph/ folder:

   • Download three.min.js from Three.js
   • Download 3d-force-graph.min.js from 3D Force Graph

4. Configure Project

   • Set development team in Signing & Capabilities
   • Enable required capabilities (see permissions below)
   • Build and run on iOS 17+ device/simulator

App Permissions

• Microphone: Voice recording for thoughts
• Speech Recognition: On-device transcription
• HealthKit (optional): Health correlation analysis
• Location (optional): Geographic context for thoughts
• Face ID/Touch ID (optional): App security lock

Data Model & Knowledge Graph

Core Data Entities

• Thought: Audio recordings with transcribed text and metadata
• User settings: Preferences and configuration

Kuzu Graph Schema

• Nodes: Thought, Keyword, Person, Emotion, Activity, Location, HealthContext
• Relationships: SIMILAR, MENTIONS, INVOLVES_PERSON, EXPRESSES_EMOTION, LOCATED_AT, HAS_HEALTH_CONTEXT
• Advanced Queries: Pattern analysis, sentiment tracking, relationship mapping

MCP Tools Available

• search_thoughts - Natural language thought search
• analyze_patterns - Emotion/keyword/activity pattern analysis
• find_similar_thoughts - Semantic similarity search
• get_emotional_timeline - Temporal emotion tracking
• get_person_interactions - Social relationship analysis
• get_health_correlated_thoughts - Health-thought correlations
• get_database_stats - Knowledge graph statistics

Oracle AI Assistant

The built-in Oracle assistant provides intelligent conversations about your thoughts:

• "What thoughts bring me happiness?" → Searches for positive emotions and related memories
• "What makes me stressed lately?" → Analyzes recent stress patterns and triggers
• "Fun memories with [Person]" → Retrieves person-specific positive experiences
• "How does my sleep affect my mood?" → Correlates health data with emotional states
• "Show me thoughts from last week" → Temporal filtering with context

Key Innovations

1. Hybrid Storage Architecture

• Core Data for reliable app data and settings
• Kuzu for complex graph relationships and advanced queries
• Vector embeddings for semantic search across both systems

2. On-Device AI Stack

• Local LFM2-1.2B inference with Leap iOS SDK (no API keys needed)
• MCP protocol for structured AI-database interactions
• Intelligent query classification and tool selection
• Multi-purpose AI: conversation, extraction, function calling, and Cypher generation

3. Privacy-First Design

• Zero external network requests for core functionality
• All AI processing happens locally on device
• Advanced encryption for sensitive thought data
• Optional biometric security layers

4. Contextual Intelligence

• Health data correlation (heart rate, sleep, activity)
• Location-based memory clustering
• Temporal pattern recognition across multiple dimensions
• Social relationship mapping and analysis

Roadmap & Future Enhancements

• Advanced NLP: Named entity recognition, topic modeling
• Visualization: More chart types, graph layout algorithms
• Export: Secure backup and data portability options
• Integrations: Calendar events, weather correlation
• Multi-modal: Photo and document attachment support

Security & Privacy

• Local-First: All data processing happens on your device
• End-to-End Privacy: No cloud sync, APIs, or external services
• Secure Storage: File-level encryption with iOS security features
• Minimal Permissions: Only requests necessary device access
• Biometric Protection: Optional Face ID/Touch ID app lock
• Transparent Processing: Open source architecture, no hidden data collection

Requirements

• Device: iPhone running iOS 17.0 or later
• Storage: ~2-3GB for app, models, and user data
• Performance: A12 Bionic chip or newer recommended for optimal LFM2-1.2B performance
• Network: Not required for core functionality (100% offline capable)

---

Your thoughts, your insights, your privacy.
Experience the future of personal knowledge management with Tapestry.