QMD — Vault Semantic Search

SkillSearch

Lets your agent search your notes vault by meaning instead of exact keywords, returning relevant snippets.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the QMD — Vault Semantic Search skill

About this capability

Search the vault using QMD semantic search. Use PROACTIVELY before reading files. Preference order: (1) MCP tools — mcp__qmd__query, mcp__qmd__get, mcp__qmd__multi_get, mcp__qmd__status — if they appear in your tool menu, use them first; (2) CLI `qmd --index <name> ...` as fallback; (3) Grep/Glob on

What this skill tells your AI

The instructions your AI receives, as published by breferrari/obsidian-mind in .claude/skills/qmd/SKILL.md and read by ahel’s review.

Before reading vault files directly, search with QMD first. It returns relevant snippets without burning context on full file reads.

Tool Surfaces — Preference Order

Pick the highest surface available and stop. All three read the same SQLite store (see Named Index), so they return the same results; the choice is about interface ergonomics and context cost.

  1. MCP tools (preferred when registered)mcp__qmd__query, mcp__qmd__get, mcp__qmd__multi_get, mcp__qmd__status. If these appear in your tool menu, the MCP server is live and pre-scoped to this vault. Call them directly — no --index argument needed, no shell involved, typed arguments. This is the intended path during a session.
  2. CLI fallbackqmd --index <name> query|search|vsearch|get|multi-get. Use when the MCP server isn't registered (older vault clones, Windows install drift, disabled in .mcp.json), or for one-off shell checks outside a session. Always pass --index <name>.
  3. Grep / Glob / Read — last resort, only when QMD is not installed at all. These burn context on full file reads and don't rank by relevance.

Never fall through surfaces without cause — if MCP is registered, calling the CLI in a Bash tool is just a slower path to the same result.

Named Index (This Vault)

This vault uses a named QMD index so queries, updates, and context strings stay scoped to this vault — not blended with any other vault that shares the machine. Every QMD command in this document passes --index <name>.

The name resolves the same way on all three surfaces (CLI, MCP, SessionStart hook), so they always point at the same SQLite store:

  1. qmd_index in vault-manifest.json, when set — an explicit pin.
  2. Otherwise the vault folder name, slugified. The template ships qmd_index empty so every install gets its own store without being asked a question.

Pin the field if you want a name that survives renaming the vault folder, or want to keep a store you already built.

Read the resolved name before running commands rather than assuming either branch:

INDEX=$(node -e '
  const ok = (s) => typeof s === "string" && /^[A-Za-z0-9][A-Za-z0-9._-]*$/.test(s);
  const m = JSON.parse(require("fs").readFileSync("vault-manifest.json", "utf8"));
  if (m.qmd_index && !ok(m.qmd_index)) {
    process.stderr.write("qmd_index is set to an invalid name: " + JSON.stringify(m.qmd_index) + "\n");
    process.exit(1);
  }
  const slug = require("path").basename(process.cwd()).normalize("NFKD").toLowerCase()
    .replace(/[^a-z0-9._-]+/g, "-").replace(/-{2,}/g, "-").replace(/^[^a-z0-9]+|[-._]+$/g, "");
  const name = [m.qmd_index, slug, m.template].find(ok);
  if (!name) { process.stderr.write("no usable qmd index name — set qmd_index in vault-manifest.json\n"); process.exit(1); }
  process.stdout.write(name);
') || exit 1
qmd --index "$INDEX" query "..."

The value is used as both a CLI argument and a filesystem path (~/.cache/qmd/<name>.sqlite), so an invalid pin must fail here rather than propagate — hence the explicit check instead of m.qmd_index || slug.

In-session, substitute the resolved value directly in your commands — it is stable across the vault's lifetime unless the folder is renamed.

Commands

Each operation lists the MCP call first, then the CLI equivalent. Prefer MCP when available.

Search (pick one per query)

  • Hybrid (best quality)mcp__qmd__query with searches=[{type:"lex", query:"..."}, {type:"vec", query:"..."}] and an intent arg. CLI: qmd --index <name> query "..." --json -n 10. Use for complex or conceptual queries.
  • Keyword (fast BM25)mcp__qmd__query with searches=[{type:"lex", query:"..."}]. CLI: qmd --index <name> search "..." --json -n 10. Use for exact terms, names, ticket numbers, dates.
  • Semantic onlymcp__qmd__query with searches=[{type:"vec", query:"..."}]. CLI: qmd --index <name> vsearch "..." --json -n 5. Use for exploratory queries where you don't know the exact words.

Retrieve

  • By pathmcp__qmd__get with path arg. CLI: qmd --index <name> get "path/to/file.md".
  • By docidmcp__qmd__get with #abc123 syntax. CLI: qmd --index <name> get "#docid".
  • Batch by globmcp__qmd__multi_get with pattern arg. CLI: qmd --index <name> multi-get "org/people/*.md" --json -l 40.

Index Management (CLI only — MCP exposes read surfaces)

  • qmd --index <name> update — Re-index after file changes (fast, ~1-2s incremental). The SessionStart hook and the PostToolUse debounced refresh both run this automatically; manual invocation is rarely needed.
  • qmd --index <name> embed — Regenerate vector embeddings (slower, run after bulk changes).

Bootstrap (Fresh Clone)

The QMD SQLite store lives outside the repo (~/.cache/qmd/<index>.sqlite), so a fresh clone starts with no index. Run the bootstrap once:

node --experimental-strip-types .scripts/qmd-bootstrap.ts

It resolves the index name (pinned qmd_index, else the vault folder slug) and reads qmd_context from vault-manifest.json, registers the collection, attaches the vault context, walks the vault, and generates embeddings. Idempotent — safe to re-run.

When to Search

Each case lists the preferred MCP call; fall back to the CLI equivalent if MCP isn't registered.

  • User mentions a past decision, incident, person, project → mcp__qmd__query (hybrid).
  • User asks "what did we decide about X" → mcp__qmd__query (hybrid with intent describing the decision context).
  • User mentions a person by name → mcp__qmd__query with searches=[{type:"lex", query:"<name>"}].
  • Before creating a new note → mcp__qmd__query with searches=[{type:"vec", query:"<topic>"}] to check for existing content.
  • After creating a note → mcp__qmd__query with searches=[{type:"vec", query:"<note title>"}] to find notes that should link to it.
  • Loading context for review prep → mcp__qmd__multi_get with pattern="perf/evidence/*.md".
  • Loading 1-on-1 context → mcp__qmd__query with searches=[{type:"lex", query:"<person name> 1-1"}].

After Bulk Changes

Run qmd --index <name> update && qmd --index <name> embed in the shell. The SessionStart hook and the PostToolUse debounced refresh run update automatically, but embed should be run explicitly after sessions that create many notes.

Signals

GitHub stars
5k
Forks
532
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
qmd-breferrari
Source
github.com/breferrari/obsidian-mind