How to set up local hybrid AI search for a 30k-note Obsidian vault with qmd and MCP
dSebastien · x · 2026-11-06
Sébastien Dubois published part 11 (32-min read) of his Obsidian search series, detailing a hands-on local AI search setup for a 30,000-note vault:
- qmd local hybrid search: BM25 keywords + embeddings + query expansion + reranker, all on-machine (2.2 GB of models). Hard-earned lesson: his index once silently missed a whole folder of 10,000 notes, so upkeep matters — run qmd status before trusting results, schedule qmd update and qmd embed, and scope collections with ignore patterns.
- Typed queries: use lex: for exact words, vec: for meaning, hyde: for a hypothetical answer; let the expansion model write them when unsure.
- MCP servers: qmd MCP for semantic search without Obsidian running; his REST/MCP server for Obsidian's native search, backlinks and Bases; the OSK plugin's server for locating notes.
- Key stance: don't hand your agent a map of the vault — make it search.
Includes full commands, config files, a refresh schedule, and a copyable AGENTS.md snippet.
Related event: Developer builds local hybrid search for 30,000 Obsidian notes with qmd(2 posts)→
More from coding & agent
- Google open-sources Elephant-Goldfish, the plugin thousands of its engineers use to rein in sloppy AI code — JafarNajafov · 2026-11-06
- Matt Pocock on why software fundamentals matter more than ever in the AI coding era — mattpocockuk · 2026-11-06
- Floci: open-source local AWS emulator runs 120+ services in 24ms, free LocalStack alternative — thisguyknowsai · 2026-11-06
- Garden Skills' motion-showreel turns Claude Code into a code-rendered video studio — thisdudelikesAI · 2026-11-06
- When AI proofreads your Chinese comments line by line, code becomes a foreign language — ducha_aiki · 2026-11-06
- Open-source social-api pulls 1,000 competitor ads for under $0.05, wired into Claude Code — ayushtweetshere · 2026-11-06