Chunk Overlap Solves the Right Problem at the Wrong Time

Thu, 23 Jul 2026 00:00:00 +0000

Introduction

I run a local-first knowledge base over my own notes: markdown files and PDFs in a folder get chunked, embedded with OpenAI, and stored in PostgreSQL with pgvector. Semantic search and a RAG chat sit on top.

Read anything about building one of these and you will be told to overlap your chunks. Ten to twenty percent is the usual number, so that an idea sitting on a chunk boundary is not cut in half. My own improvement notes for the project say exactly this, filed under “Impact: High”.

Retrieval on Jamie Craane personal blog

Chunk Overlap Solves the Right Problem at the Wrong Time

Introduction