Documentation
Advanced config
advanced.yaml – embeddings, search, clustering, AI timeouts, and prompts for power users.
Everything in Settings has a sensible default. For the knobs that don’t
belong in a Settings window, there is advanced.yaml:
~/Library/Containers/com.gisti.app/Data/Library/Application Support/com.gisti.app/advanced.yamlSettings → Storage → Location opens that folder directly in Finder – easier than typing the path by hand. Gisti writes the file itself on first launch, with every key present and commented; edit the ones you want and restart. Keys added by later versions take their defaults, so an old file keeps working.
Sections
embedding:
modelId: mlx-community/Qwen3-Embedding-0.6B-4bit-DWQ
queryInstruction: "Retrieve relevant documents and notes matching the search query"
maxTokenLength: 1024
maxBatchSize: 4
gpuCacheLimitMb: 512
maxTextLength: 4000
indexBatchSize: 8
indexBatchSleepMs: 200
textChunkPreviewLength: 200
search:
minSimilarity: 0.4 # semantic match threshold
topK: 20 # candidates per search
similarTopK: 5 # candidates for the "Similar" accordion
similarMinSimilarity: 0.55
topicsTopK: 50 # candidates feeding Topics clustering
rrfK: 60 # Reciprocal Rank Fusion constant (hybrid search)
debounceMs: 250
itemTextCap: 16384 # per-item text length fed to the index
savedQueries: [] # queries offered in the ⌘F menu; see Stream → Saved searches
fuzzy:
minWordLength: 4 # shorter words match exactly only
maxEditDistance: 1 # typos tolerated per word
exactMatchSkipThreshold: 20
maxResults: 5
clustering:
maxIterations: 30 # KMeans iteration cap
maxTopics: 60 # runaway guard on the initial k; the real count is sqrt(N/2)
minVectors: 4 # elements needed to form a topic
minItemChars: 4 # shorter records skip clustering and go to Unsorted
labelCandidateCount: 100 # items sampled per cluster when naming it
labelTextScanLimit: 300 # characters of item text read when its title is generic
labelMinItems: 2 # items of a topic a word must appear in to name it
labelMaxTokenLength: 20 # longest token that can be a name (a base64 run isn't a word)
labelMinScore: 0.07 # naming threshold, 0..1 – see below
minNamedTopics: 8 # give at least this many a single-word name
maxClusterShare: 0.06 # topics larger than this share (but never smaller than
# two average topics) get split in two
aiTimeouts:
summarizeSeconds: 60
chatSeconds: 120
titleSeconds: 30
testSeconds: 10
aiLimits:
summarizeInputMaxChars: 10000
summarizeMaxTokens: 8192
chatMaxTokens: 4096
titleMaxTokens: 32
attachmentTextMaxChars: 20000
attachmentImageMaxDimension: 1568
askAIContextMaxChars: 40000 # text budget for the whole Ask AI request
askAIContextMaxImages: 8 # images per Ask AI request, counted separately
attachmentMaxFileBytes: 10485760
attachmentImageMaxFileBytes: 52428800
digestInputMaxChars: 20000 # char budget for the Gist-mode digest prompt
digestItemsPerTypeMax: 50 # per-type item cap when sampling for the digest
clipboard:
pollIntervalSeconds: 0.2
maxTextLength: 5000000 # 0 disables the limit; longer copies are skipped, not truncated
ignorePasteboardTypes: # copies carrying one of these types are never recorded
- com.agilebits.onepassword
- net.antelle.keeweb
- de.petermaurer.TransientPasteboardType
- "Pasteboard generator type"
- com.typeit4me.clipping
ignoreRegexps: [] # text matching any of these is never recorded
maxRichTextBytes: 131072 # largest RTF/HTML kept beside a clip; 0 stores plain text only
peek:
holdDelaySeconds: 0.3 # Force Touch hold before the preview appears
fadeOutSeconds: 0.15
pressureOpacityBase: 0.3 # opacity ramp while building up Force Touch pressure
pressureOpacityGain: 0.58
snap:
topOffset: 150 # px from the top edge when snapping the window
ui:
showWhitespaceMarker: true # flag clips that start or end with a space/newline
contentViews:
detectMaxBytes: 5000000 # longest record offered a format chip
tableMaxRows: 1000 # rows of a CSV/TSV table shown at once; copying is never capped
markdown:
loadRemoteImages: true # false: remote images in markdown stay placeholders, nothing is fetched
syntaxHighlight:
maxBytes: 262144 # longest text that gets syntax colors
maxBlockBytes: 65536 # longest single ``` block that gets them
detectCodeNotes: true # open a note that is code with colors and a language chip
codeNoteMaxBytes: 32768 # longest record that guess is attempted on
codeNoteMinLines: 5 # shorter than this is a command or a sentence, not a file
codeNoteMinRelevance: 8 # floor under the winning language's score
codeNoteRelevanceGap: 4 # how far ahead of the runner-up it must finish
codeNoteRelevanceGapFromEditor: 2 # …when the clip came from an editor or terminal
codeNoteRelevanceGapShare: 0.15 # the same lead as a share of the winner's own score
codeNoteRelevanceGapShareFromEditor: 0.08
prompts:
summarization: >
You are a concise summarization assistant. ...
dailyDigest: >
Create a digest based on the following notes, clipboard items, ...
digestSystem: >
You are a digest assistant that summarizes a user's personal activity ...
chatSystem: >
You are a helpful AI assistant. Answer questions, discuss ideas, ...
titleGeneration: >
Generate a very short title (3-6 words) for a chat ...| Section | What it controls |
|---|---|
embedding |
MLX model id, query instruction, token/batch limits, GPU cache, indexing batches |
search |
semantic thresholds, top-K, similar-items and Topics candidate counts, hybrid ranking (RRF), debounce, per-item text cap, saved searches |
fuzzy |
typo tolerance and result caps for text search |
clustering |
Topics: max topics, KMeans iterations, and how topics get named |
aiTimeouts |
per-request-type timeouts in seconds |
aiLimits |
input/output size caps: summarize/chat/title token limits, attachment text/image/file caps, Ask AI context budget, digest sampling budget |
clipboard |
clipboard poll interval, max copied-text length, ignore rules by pasteboard type and by content, ceiling on kept formatting |
peek |
Force Touch preview: hold delay, fade-out, and pressure-opacity curve |
snap |
window snapping: top-edge offset |
ui |
whether a stream row flags leading/trailing whitespace in the clip or note it shows |
contentViews |
Formatted views: largest record offered a chip, rows of a table shown at once |
markdown |
whether rendered markdown may fetch the remote images it references |
syntaxHighlight |
syntax colors: size ceilings, and how reluctant the app is to call a note code |
prompts |
system prompts: summarization, daily digest, digest system prompt, chat, title generation |
Tuning how topics are named
labelMinScore is the one knob worth touching if topic names don’t feel
right. It runs on a 0…1 scale, where 1 means “this word is in every item of
the topic and in no other topic”. The default 0.07 reads as the word
appears in about 7% of the topic’s items and is unique to it.
The threshold never removes a topic – it only decides whether the card gets
a single word or a two-word name like Docker · compose.
- Too many two-word names – lower it (0.04–0.05). More topics get a single-word name, some of them weak.
- Names look arbitrary – raise it (0.12–0.2). Fewer single words, more two-word labels, which say more anyway.
minNamedTopics is the floor: if fewer topics than that clear the
threshold, the best of the rest get a single-word name anyway.
labelMinItems is the other useful knob. A word one item happens to
contain isn’t what a topic is about, and letting it compete also drags down
the score of every word that is characteristic. Raise it if names still
look like they came from whatever you copied last.