LLM Wiki — Synaptic Lattice

qmd

Properties1
tagsentity, repository, tool, llm-wiki, search, retrieval

Leitura de 2 min

Overview

qmd (by Tobias Lütke) is a hybrid BM25 + vector search engine purpose-built for markdown wikis. It runs a three-model local pipeline (query expansion + reranker) with no LLM API calls. In the comprehensive report it is the recommended search layer when a wiki outgrows index.md grep (Pattern B / Scale tier 100–500 pages).

Key Facts

  • Type: hybrid search engine (BM25 + vector) for markdown
  • Pipeline: three local models — query expansion, then reranker
  • Cost: no LLM API calls; runs locally
  • Role: search MCP (qmd-mcp) for scaling past ~500 pages
  • Author: Tobias Lütke (Shopify founder)
  • Released May 2026 (per the comprehensive report timeline)

Features

Core Architecture

  • Hybrid search — BM25 keyword + vector semantic search combined
  • Three-model local pipeline — query expansion + reranker, all local
  • Zero API calls — runs entirely offline, no LLM API costs
  • MCP server — qmd-mcp exposes search to agents

Search & Retrieval

  • BM25 + vector — best of keyword and semantic worlds
  • Query expansion — local model rewrites queries for better recall
  • Reranker — local model re-ranks results for precision
  • Markdown-native — built specifically for markdown wiki content

MCP / Tools

  • qmd-mcp — Model Context Protocol server for agent integration
  • CLI tool — command-line search interface
  • Local-first — no cloud dependency

Integration

  • Agent-compatible — works with any MCP-capable agent
  • Obsidian compatible — searches Obsidian vaults directly
  • Wiki-native — understands wikilink structure

Data & Storage

  • File-based index — no database required
  • Markdown-native — indexes .md files directly
  • Privacy-preserving — all data stays local

Relationships

  • llm-wiki — (type: enables) search backend for growing wikis
  • agent-memory-systems — (type: belongs-to) retrieval layer for the artifact-first family
  • cognee — (type: sibling-of) alternative scale path (graph vs search); the report offers both
  • model-context-protocol — (type: uses) exposes qmd-mcp hybrid search
  • vault-roadmap — (type: belongs-to) the vault’s own scale phase anticipates local search like this

Implications

qmd addresses the vault’s own “what happens when index.md breaks” question. It is the lightweight, local, privacy-preserving search option — directly relevant to the user’s earlier concern about the vault being “too manual”: local search removes the need to grep by hand without sending data to an API.

Sources

Visão de gráfico

Backlinks