Week 4 🍏

Local RAG & Personal Knowledge Bases

Target: Memory and Retrieval

🎯 Objective

Learn the most important pattern in modern AI: Retrieval-Augmented Generation (RAG). You will build a system that allows an LLM to "read" your app's private database to answer questions.

Daily Breakdown

Day 1

Introduction to Embeddings

Topic: Turning text into math.

Tasks:
  • Use NLModel or a Core ML embedding model to convert sentences into vectors (arrays of Floats).
  • Calculate Cosine Similarity between two vectors in Swift.
Day 2

The Vector Database (SwiftData + Embeddings)

Topic: Local Storage for AI.

Tasks:
  • Extend SwiftData to store vector embeddings alongside traditional metadata.
  • Implement a basic "Nearest Neighbor" search logic to find relevant data chunks.
Day 3

Chunking & Context Injection

Topic: Managing the "Context Window."

Tasks:
  • Create a "Chunking Engine" that breaks long PDFs/Notes into 500-token pieces.
  • Logic: User Query → Search Vector DB → Get Top 3 Chunks → Inject into LLM Prompt.
Day 4

Ranking & Re-ranking

Topic: Improving Accuracy.

Tasks:
  • Implement a "Re-ranker" logic to filter out irrelevant search results before they hit the LLM.
  • Use Natural Language Framework to tag parts of speech to improve search intent.
Day 5

End-to-End RAG Architecture

Topic: System Design.

Tasks:
  • Optimize the pipeline: Ensure the "Retrieval" phase doesn't block the UI thread.
  • Implement "Source Attribution" (showing the user exactly which document the AI used for its answer).
🧪

Friday Lab: "Second Brain"

Project: A private "Personal Wiki" app.

  1. Feature: User can paste long articles into the app.
  2. Feature: The app automatically embeds and stores them.
  3. Feature: A chat interface where you ask "What did I learn about [Topic] last week?" and the app answers using only your stored notes.
  4. Constraint: Zero network calls. Everything must reside in the local .sqlite store.