nodecode-md-search

Markdown memory: full-text search over your own markdown folders, ranked by bm25. A Nodecode rewrite of kuehnberger/hermes-qdrant-memory by kuehnberger.

The model gets one tool, md_search, that answers the passages of your notes, vault or documentation that match a query: the file, the heading chain it sits under (Guide > Install > Linux) and one excerpt around the match. /md-search runs the same search by hand and shows what the index holds.

Nothing is embedded and nothing leaves the machine. The index is built in the Nodecode process from the folders you name, chunked by heading, and ranked with bm25 — no SQLite, no model download, no Qdrant, no network.

Get it

Tell the agent "add the md-search cell", or run nodecode add nc://md-search.

Set it up

Name the folders to search. Each entry is a path, or label=path when the label an answer quotes should be something other than the folder's own name:

// ~/.nodecode/config.jsonc
{
  "md-search": {
    "roots": ["~/Documents/notes", "vault=~/Documents/vault", "~/src/project/docs"],
    "refresh_seconds": 300
  }
}

Member	Default	What it does
`roots`	none	the folders searched, each a path or `label=path`; a directory starting with a dot is skipped
`skip`	`.git` `.obsidian` `.trash` `node_modules` `__pycache__` `.venv` `venv`	directory names never walked
`max_results`	`5`	how many passages one search answers when the caller names no limit
`snippet_chars`	`170`	how many characters of a passage one result quotes
`max_chunk_chars`	`1400`	the longest passage before a section is split on its paragraphs
`max_file_kb`	`5120`	a markdown file bigger than this is not indexed
`max_files`	`5000`	the most markdown files one index holds
`refresh_seconds`	`300`	how long an index is reused before the folders are walked again; `0` walks them on every search

With no roots the tool still registers and names the setting it needs, so "set up md-search" finishes it in the conversation, in /settings, or in the page's Settings tab.

Use it

Ask for what you know you wrote — "what did I write about the backup rotation?" — and the model calls md_search. Search by the words you wrote, not by a question phrased for a model: the ranking is bm25 over the passage's terms, with a section heading counted twice so a title outranks a mention in a body.

/md-search                 the corpus and what the index holds
/md-search backup rotation  the same search, by hand
/md-search reindex          rebuild the index now

A note edited since the index was built is found again after refresh_seconds, or at once with /md-search reindex. A finished turn does not write anything: this cell reads, it never edits a note.

Test

.github/test-cell.sh tools/nodecode-md-search