Ditching the Vector DB: Using an INDEX.md File for Agent Retrieval, and Its Tradeoffs

Low_Rush_8535 · reddit · 2026-10-10

For a collection of a few dozen docs, the author skipped embeddings and vector databases entirely, using an INDEX.md file instead: one line per document with a short description. A doc-searcher subagent reads the index, picks relevant files, and reads them; a separate summarizer generates descriptions, which are regenerated when a document changes.

The real win is context isolation: the parent agent receives only the subagent's final result, not every intermediate read. But the author is candid that this isn't free retrieval—the subagent still reads the index and selected docs, and if the parent needs the original material, passing file paths back costs tokens too. Context isolation and total token savings are different things; with no clean benchmark, the author declines to claim a savings percentage.

Boundaries: for small collections it means fewer moving parts; at thousands of docs the index itself becomes a lot to read, and a vague description can hide a relevant file.

Original post →

Ditching the Vector DB: Using an INDEX.md File for Agent Retrieval, and Its Tradeoffs

More from coding & agent