GNO vs Qmd

The LLM Wiki search layer

Karpathy’s LLM Wiki pattern needs a local search engine over markdown files with hybrid BM25 + vector retrieval and an MCP server so agents can query the compiled knowledge. Both Qmd and GNO fit this role, and both run the whole pipeline on the machine with GGUF models through node-llama-cpp. The differences are the corpus each one reads, the tool surface each one advertises, and what sits beyond search.

Format support

• Qmd — Text files matched by a per-collection glob (**/*.md by default); code in TypeScript, JavaScript, Python, Go, and Rust gets AST-aware chunking. PDFs and Office files are outside its indexer.
• GNO — Markdown, PDF, DOCX, XLSX, PPTX, plain text, and code files, plus mail, calendar, transcript, JSONL, and browser exports; indexed-document language detection covers seven languages

Retrieval

• Qmd — search is BM25, vsearch is vector, and query is the hybrid path: FTS + vector + query expansion + LLM reranking, fused with RRF. The generate model role is a fine-tuned query-expansion model; Qmd returns ranked documents and qmd:// URIs rather than generated answers.
• GNO — BM25 + vector + cross-encoder reranking with query expansion, deterministic Context Capsules with exact line spans and content hashes, opt-in verified Ask that abstains below full claim support, and gno query diagnose for the document that should have come back

MCP surface

Qmd’s MCP server exposes four read tools: query, get, multi_get, and status, over stdio or a loopback HTTP transport that checks Origin and Host. The corpus is written from the CLI (collection add, update, embed, context), so its MCP surface is read-only by construction. Since 2.8.3 it speaks MCP revision 2026-07-28 through the official SDK 2.x while still serving 2025-era stdio clients.

GNO advertises 38 read-only tools by default and adds 21 write tools plus a read-only request-status lookup behind --enable-write, for 60 total; the opt-in core profile advertises 7 read tools (9 with writes) for sessions that should pay for fewer tool definitions. Both profiles speak the 2025-11-25 and 2026-07-28 protocol revisions on stdio and the resident HTTP endpoint. Beyond MCP, one command installs the agent skill for Claude Code, Codex, OpenCode, OpenClaw, or Hermes, and a loopback REST API and TypeScript SDK serve custom tooling.

Memory

Qmd is a retrieval layer: what an agent remembers is whatever lives in the files it indexes. GNO adds a memory contract on top of the same index. gno remember stores one fact as a markdown file under explicit scopes, a change is a hash-checked supersede that keeps the predecessor on disk, and gno recall returns the current facts, at most 8 under 512 tokens, cited by gno:// URI. The contract runs on the CLI, MCP (gno_recall read-only, gno_remember behind the write flag), REST, and the SDK.

Beyond search

GNO is also a workspace: a browsable collection tree, a knowledge graph, safe editing, and optional hosted publishing via gno.sh. Qmd stays a CLI and MCP server by design, a complete tool when search is the whole job.

Qmd claims on this page were checked against the README on tobi/qmd main (commit dbfd0b4, 2026-08-18) and the 2.8.3 changelog (2026-08-16).