Notion RAG / Second Brain

A retrieval system that transforms a large personal knowledge base into searchable, contextual AI memory.

Ingestion → Chunking → Embeddings → Retrieval → LLM Reasoning

01 — Overview

This project was an early technical exploration of Retrieval-Augmented Generation (RAG) applied to a personal Notion workspace (a "Second Brain"). It laid the groundwork for later concepts around personal intelligence and evidence-backed AI.

02 — Problem

Traditional note-taking creates information accumulation but does not necessarily create usable knowledge. Thousands of Notion pages are useless if you cannot instantly recall the specific context you need. The goal was to ask questions and get answers grounded entirely in my own notes.

03 — Why it was technically difficult

Naive RAG (just chunking text and using cosine similarity) produces terrible results on structured personal notes. The system needed to handle deeply nested hierarchies, maintain document metadata, and retrieve contextually relevant chunks without losing the surrounding meaning.

04 — Architecture

Notion pages → ingestion → parsing → chunking → metadata → embeddings → vector retrieval → contextual retrieval → LLM reasoning → answer with source context

05 — Key engineering decisions

• Semantic chunking: Instead of splitting by character count, the pipeline parsed Notion blocks to keep semantic units (paragraphs, lists, code blocks) intact.
• Rich metadata injection: Every chunk was enriched with the parent page title, tags, and date to give the embedding model context.
• Incremental indexing: Implemented a sync engine that only re-embedded pages that had changed since the last run, saving API costs and time.

06 — AI architecture

User Query → Query Embedding → Vector Search (Top K) → Context Assembly → LLM Prompting (Strict Grounding) → Response with Citations

07 — Product decisions

• Strict source attribution: The LLM was prompted to refuse to answer if the information wasn't in the retrieved context, and it was required to cite the specific Notion page for every claim.

08 — Outcome

I built a highly effective personal search engine that could accurately synthesize information across years of notes. More importantly, this project was the conceptual bridge: asking how can personal information become useful intelligence instead of remaining trapped inside documents?

09 — What I learned

I learned the practical limits of vector search and the absolute necessity of good metadata. This project heavily influenced my later work on Sovereign Intelligence, proving that reliable AI requires immaculate data pipelines.