Settings reference

Every key numen.json can hold. The pages under Settings say which of these are worth touching and why; this one is here so that nothing is missing.

Each heading below is a section of the file, and the keys under it are written the way they are written inside that section — model.name under indexing.embedding is indexing.embedding.model.name if you write it out in full.

A key you leave out keeps its default, so a file naming one setting is a complete file. A key numen does not know is carried through untouched. A row with no kind beside it is a group holding the keys under it.

Two things are not settings and are not in here: which files count as notes, and what a vault’s own folder is called. Both are asked for on the command line, by numen-cli.

v	a number	the shape of the file. Nothing reads it yet, and it is written so that the day a section changes shape there is something to tell the two apart.
plugins		the configuration for modular plugins, keyed by plugin ID.

appearance

How the window is drawn.

interface_scale	a number	how large the window is drawn: its chrome, its controls, the spacing between them and the type in them.
text_scale	a number	how large the text a person reads is set: a note, a book, an answer, the editor.
mode	text	which half of a colour pair the window takes: system, light or dark.
theme	text	the stylesheet the window wears, named by its origin and its filename: preset:dracula ships here, mine:dracula is the person’s file.
nest_headings_under_notes	yes or no	whether a node in the plex nests the headings of its note under the box.
nested_heading_limit	a number	how many of those headings stand under a node at once, the rest being reached by scrolling.

indexing

How a vault is made searchable.

embedding		which model turns text into vectors, and how it is reached.
embedding.model		what a vector is: everything that decides the space it lands in.
embedding.model.name	text
embedding.model.dimensions	a number
embedding.model.max_tokens	a number
embedding.model.pooling	text
embedding.indexing		where a vector is made: on this machine, or by a service.
embedding.indexing.use	text
embedding.query		where a vector is made: on this machine, or by a service.
embedding.query.use	text
embedding.floor	a number
recognition		how a scanned document is read when a person asks for it.
recognition.runtime	text	the shared library of ONNX runtime. Empty means the one beside the application.
recognition.dir	text	a folder holding the models. Empty means the folder beside the application, and then the download cache.
recognition.download	yes or no	allows fetching what is not on this machine.
recognition.layout		divides a page into its parts and puts them in reading order.
recognition.layout.name	text	where the model is fetched from, and path is a file on this machine.
recognition.layout.path	text	this model as a file on this machine.
recognition.layout.labels	a list of words	the parts the model knows, in the order of its class ids.
recognition.layout.minimum	a number	the score a part carries to be a part at all.
recognition.layout.overlap	a number	how much of two parts may be common before the second is taken to be the first found again.
recognition.layout.margin	a number	what a part is widened by before it is read, because the first letter of a line sits on the boundary the model drew.
recognition.detect		finds the lines one part holds.
recognition.detect.name	text	where this model is fetched from.
recognition.detect.path	text	this model as a file on this machine.
recognition.detect.max_side	a number	the longest side a part is read at.
recognition.detect.expand	a number	how many pixels a found line is widened by, in the image the detector reads: the part scaled so that its longest side is max_side.
recognition.detect.minimum	a number	the heat a pixel carries to be part of a line.
recognition.recognise		reads what a line says: which model, at what size a page and a line reach it, and how much of the machine it takes.
recognition.recognise.name	text	where this model is fetched from.
recognition.recognise.path	text	this model as a file on this machine.
recognition.recognise.dict	text	the model’s characters, one to a line. Empty is the ordinary case: the model carries them, and asking it is the only way to be sure.
recognition.recognise.classes	a number	how many characters the model knows and two more.
recognition.recognise.dpi	a number	what a page is rendered at.
recognition.recognise.height	a number	what a line is scaled to before it is read.
recognition.recognise.jobs	a number	how many recognition sessions run at once. Zero takes the performance profile.
recognition.recognise.threads	a number	how many threads one model may use. Zero takes the performance profile.
recognition.regions		says what the parts of a page are for. A part the model names that body does not carry is not read.
recognition.regions.body	a list of words	carry what the document says.
recognition.regions.head	a list of words	open a part of the document, outermost first: where a name stands in the list is how deep the part it opens sits.
recognition.proofread		names the profile a recognition is put right at.
recognition.proofread.with	text	the profile, by the name the profiles carry it under.
recognition.proofread.automatically	yes or no	whether a text already written down is put right without anybody asking for it.
proofreading		what puts a recognition right. Naming no profile here is naming no proofreader, and a recognition is used as it was read.
proofreading.max_edit_distance	a number	how far a correction may stand from the line as read and still be a correction: the Levenshtein distance between their letters, as a share of the longer of the two.
proofreading.profiles		the places a text is put right at, by the name a consumer asks for one under.
transcription		how a recording is listened to: which models hear it, where they came from, and how the speech in it is found.
transcription.runtime	text	the ONNX shared library. Empty means the one beside the application, and then the one the platform holds.
transcription.dir	text	a folder holding the models. Empty means the folder beside the application, and then the download cache.
transcription.download	yes or no	allows fetching what is not on this machine.
transcription.threads	a number	how many threads the encoder may use. Zero takes one per core, up to eight.
transcription.model		turns speech into words. It is exported as three graphs and the pieces they write, and all four are one model: three graphs from two exports answer with nothing anybody can read.
transcription.model.name	text	what this model is called in the record kept beside a text.
transcription.model.from	text	the folder the four files are fetched from.
transcription.model.encoder	text	encoder, decoder, joiner and tokens are the files on this machine.
transcription.model.decoder	text
transcription.model.joiner	text
transcription.model.tokens	text
transcription.segmenter		finds where in a recording somebody is speaking.
transcription.segmenter.name	text	what this segmenter is called in the record kept beside a text.
transcription.segmenter.from	text	where the model is fetched from, and path is a file on this machine.
transcription.segmenter.path	text
transcription.segmenter.threshold	a number	how sure the model has to be that a window carries speech.
transcription.segmenter.silence	a number	how much quiet, in milliseconds, closes a segment.
transcription.segmenter.pad	a number	how many milliseconds are kept on each side of a segment, so that the first and last sound of a word are inside it.
transcription.segmenter.longest	a number	how many milliseconds one segment may run to.
transcription.segmenter.shortest	a number	how many milliseconds a segment carries to be a segment at all.
transcription.segmenter.least	a number	how many milliseconds a segment runs to before it stands as a line of its own.
transcription.proofread		names the profile a transcript is put right at.
transcription.proofread.with	text	the profile, by the name the profiles carry it under.
transcription.proofread.automatically	yes or no	whether a text already written down is put right without anybody asking for it.
transcribe_recordings	yes or no	whether a recording the vault holds no transcript for is listened to without anybody asking.
transcribe_under_mb	a number	how large a recording may be and still be listened to without anybody asking, in megabytes.

agent

Which agent answers in the panel, and what it may reach.

use	text	names the agent. Empty answers with none, and the panel says so.
serve_tools	yes or no	puts the tools on a port, which is how an agent a person runs themselves reaches this vault.
claude		code, reached by starting it and reading what it prints.
claude.command	a list of words	starts it: the command line’s path, and anything it is started through.
claude.model	text	which of its models answers — opus, sonnet, haiku, or a full name.
claude.max_steps	a number	how many times it may go to the model before it is stopped.
claude.reads_hooks_and_skills	yes or no	lets it read what is configured for it on this machine: hooks, skills, standing instructions in CLAUDE.
antigravity		google antigravity, reached over ACP protocol.
antigravity.command	a list of words
antigravity.model	text
antigravity.max_steps	a number
antigravity.reads_hooks_and_skills	yes or no
codex		openAI codex, reached over ACP protocol.
codex.command	a list of words
codex.model	text
codex.max_steps	a number
codex.reads_hooks_and_skills	yes or no

importing

How an address a link note points at is reached, and where the tools that reach it are.

captions	a list of words	the languages published words are preferred in, best first.
automatic_captions	yes or no	whether words a machine wrote count where a person published none.
copy_max_size_mb	a number	how large a copy may be. Above it, a copy asked for says the size it was refused at and nothing is fetched.
copies_to_vault	yes or no	whether a copy is kept beside the note as a file of the person’s own.
yt_dlp		what is run to reach a video. Empty asks the path.
yt_dlp.command	a list of words	what is run, so a machine that writes the path afresh at every build names whatever does know where the tool is.
yt_dlp.arguments	a list of words	handed to every run before its own: what answers for a person at a site that refuses an unattended fetch — the cookies of a browser, a token, a proxy — is that machine’s and is passed through as it stands.
yt_dlp.environment		set on every run, over what this process was started with.
ffmpeg		what brings a container to what a transcriber opens.
ffmpeg.command	a list of words	what is run, so a machine that writes the path afresh at every build names whatever does know where the tool is.
ffmpeg.arguments	a list of words	handed to every run before its own: what answers for a person at a site that refuses an unattended fetch — the cookies of a browser, a token, a proxy — is that machine’s and is passed through as it stands.
ffmpeg.environment		set on every run, over what this process was started with.

naming

How a note’s title and the name of its file are held together.

sync_title_and_filename	yes or no	whether renaming either of the two brings the other into line.

review

What a day of review is, on this person’s clock.

day_start_time	text	the hour a day of review begins at, on the clock on the wall, written as hours and minutes.

performance

The workload and resource calibration profile.

profile	text	the active preset.
custom		holds user-defined overrides when profile is ProfileCustom.
custom.embedding_batch_size	a number
custom.ocr_threads	a number
custom.ocr_sessions	a number
custom.llm_concurrency	a number