Developer Tests Four RAG Retrieval Methods on 312-Doc Knowledge Base, Only One Holds Up

A developer built a personal knowledge database called context-forge, containing 312 documents including tweets, research abstracts, blog posts, and YouTube transcripts, managed daily using Claude Code via a CLI tool. The system uses a SQLite index with markdown files and an automated pipeline that fetches, summarizes, scores, and categorizes each new document from a URL. To find the best retrieval strategy, the developer ran four different approaches over two-week trials each against real daily tasks such as writing, research, and decision-making. Vector embedding search degraded around 200 documents because topically dense, short docs produced near-identical similarity scores, while keyword-based BM25 search via SQLite FTS5 later broke down due to ambiguous query terms returning unrelated results. Of the four strategies tested, three failed in distinct and reproducible ways that the developer notes are commonly recommended in mainstream tutorials.