Filter

Hide nodes that don’t match. Filtered-out nodes and their edges disappear — not dim, actually gone.

Filtering is how you reduce visual noise to what matters. Every property on every node is filterable, with a UI tailored to its type. Multiple filters stack with AND semantics. The result is a live view that updates as you adjust controls, with edges from/to hidden nodes culled automatically.

The difference from Search matters: filter carves the view (hidden means gone), search highlights within it (non-matches dim, stay in place). You typically filter first to carve, then search to find.

Per-type UI

Knotviz gives each property type a dedicated control, because filtering a numeric range is fundamentally different from picking categorical values.

Property type	Control
number	Dual-handle range slider with min/max inputs. Histogram overlay + three scale modes (see below).
string / string[]	Searchable multi-select with chip display. Select-all / clear-all / reset shortcuts.
boolean	True / False radio. Disable the filter to include both.
date	Dual-handle range slider with ISO-formatted endpoints. Histogram overlay + three scale modes (see below).

Scale modes (number and date only)

Number and date filter panels have a small chevron next to the histogram toggle. Click it to reveal a three-way segmented control: lin · log · %. The three modes are mutually exclusive and the selection persists while you drag; the slider’s min/max selection is preserved across mode changes so toggling modes never loses your filter bounds.

Mode	Number filter	Date filter
lin (linear)	Slider position is proportional to value. Default.	Slider position is proportional to date. Default.
log	Log-scale slider + histogram; gives the low end more resolution. Requires min ≥ 0.	Inverted log: gives the recent end more resolution (matches typical clustering-toward-now). Requires dates ≥ 1970-01-01.
% (percentile)	Slider runs from p0 to p100. A drag to [10, 90] keeps the middle 80% of nodes by value. Tiny p{N} tags above the inputs show what percentile each handle is at.	Same percentile semantics on the date distribution.

Each filter has its own enable checkbox at the top-left. Disabling keeps your configured values but stops the filter from participating in the match. Handy for A/B comparisons — flip one filter on and off to see its effect in isolation.

AND across all filters

Multiple enabled filters combine with AND — a node is visible only if it matches every one of them. The match count at the top of the panel (714/1,000 nodes match) updates live as you adjust controls.

For concrete intuition: imagine filtering to age ≥ 30 AND tags includes "engineer" AND joined > 2020-01-01. A 35-year-old engineer who joined in 2022 passes all three and stays visible. A 40-year-old designer with the right join date fails the tags filter and disappears, along with every edge attached to them.

There’s no OR / NOT / XOR. Keep filters simple; combine with Search or multiple comparison passes for more complex expressions.

Hide, don’t dim

Filtered-out nodes are gone, not dimmed. Every edge attached to a hidden node is also hidden; you never see “floating” edges with a missing endpoint.

That’s the deliberate split from search: filters carve the view, search dims within what’s visible. If you want to keep context and just emphasise a subset, use Search instead.

The cost of carving is you lose context: a neighbour of a visible node might be hidden, making that node look less connected than it is. Toggling the filter off and back on is cheap — use that to flip between zoomed-in and zoomed-out views of the same data.

The productive loop: filter + colour

Filter and Analyze compose well. A typical pattern:

1. Colour by a categorical property — communities, types, risk tiers. You can now see which clusters dominate.
2. Filter by a numeric or boolean property — activity > threshold, verified = true. The coloured clusters thin out; which colour survives?
3. Add a second filter to sharpen the question. Does the high-risk cluster mostly share one community?

Stats update live with the filter set, so the distribution you see in the Statistics panel reflects only what’s currently visible — concrete enough to notice patterns you’d miss in raw tables.

Per-filter shortcuts

• Select all / Clear all on string and string[] filters — skip the multi-click when you want everything except a handful.
• Reset per filter — returns to default (disabled, full range). Doesn’t remove the filter; you can re-enable later.
• Clear all at the top — disables every filter at once. Equivalent to starting over without losing your configured values.

Gotchas

• Edges to hidden nodes disappear even if the edge itself has nothing filtered. Intentional; the alternative (orphan edges) looks broken.
• log mode only unlocks when the property’s min value is in range — numbers must be ≥ 0, dates must be ≥ 1970-01-01. Out-of-range data leaves the log button disabled (dimmed); stick with lin or pre-filter to the valid range.
• Date log is inverted vs numeric — it spreads out recent dates rather than early ones. Reason: most real-world date properties (signups, events, commits) cluster toward “now”, so fine-grained control belongs at the right end of the slider. If your data is the unusual shape — lots of old dates, few recent — lin will actually give you better resolution.
• % mode works on the dataset’s own distribution, not a global reference. If you enable a filter, narrow via other filters, then open a % slider, the percentiles still reflect the original dataset, not the currently-visible subset. (Precomputed once at load time to keep slider drags O(1).)
• string[] filtering uses any-match semantics — a node with tags [a, b, c] passes a filter that selects a alone. For must-match-all-tags, add one filter per required tag.
• The filter panel computes distinct values live from visible nodes, so disabling one filter may change the options available in another (the multi-select updates to reflect the now-visible set).