QMD

This is the concept note. For the OSS project factsheet: qmd (OSS).

QMD (Query Markup Documents, by tobi — Shopify founder Tobias Lütke) is an
on-device hybrid search engine for markdown knowledge bases: **BM25 full-text

• vector semantic search + LLM re-ranking**, all local. One binary indexes a
  folder of .md files and answers natural-language queries with cited chunks.

Why it matters here

The digital garden vault is both a public wiki (garden.h) and the LLM wiki
backing Hermes agents (How this garden works). QMD is its search layer:

• qmd update && qmd embed runs after every tending session — hybrid index
  stays in sync with the vault.
• Registered as a native MCP server (mcp_servers.qmd → qmd mcp) in Hermes
  config, so agents query it directly alongside file tools.
• Distinct from RAG-with-per-query-ingest: the wiki compiles knowledge once;
  QMD retrieves from it cheaply and with citations.

Related homelab efforts (overlap, not dependency)

The homelab local-LM roadmap (~/projects/homelab/docs/local-lm-roadmap.md §2)
plans a RAG pipeline over the same kind of markdown KB — embedding model
comparison, vector DB choice (Qdrant/LanceDB/pgvector), chunking questions.
QMD sidesteps most of that: it bundles BM25 + embeddings (GGUF via
node-llama-cpp) + re-ranking in one local binary with zero infrastructure.

Separately evaluated memory systems — Hindsight (ONNX vs PyTorch reranker
decision), agentmemory (BM25 + local embeddings, no-LLM mode) — solve a
different problem (agent memory/observations), but their hybrid-retrieval
stacks are the same architecture QMD ships as a single tool. Where those
projects planned “embedding model + vector DB” as open questions, QMD
already made those choices.

Alternatives landscape

• Dedicated vector DBs (Qdrant, LanceDB, pgvector) — infra QMD doesn’t need.
• Agent-memory servers (Hindsight, Supermemory, agentmemory) — same retrieval
  stack, different problem domain.
• Obsidian’s own plugin ecosystem (Omnisearch etc.) — GUI-bound, not
  agent-callable.

Cross-refs: qmd (OSS), Digital gardens, How this garden works,
Quartz modifications. Homelab overlap: Local LM roadmap §2
(RAG/embedding-model plans QMD obviates for the garden vault).