Knowledge base

Tab

Moves through settings, Add source, the view switch, each Retry or Remove button and browse in Sources; and through the query field, Search, the Rerank switch and each result in Test retrieval.

Arrow leftorArrow right

On the view switch, moves between Sources and Test retrieval. On a settings slider, changes the value by one step; Shift moves a larger step.

Enter

In the query field, runs the search. On a suggestion, runs that question. On a result, opens or closes the chunk in context.

Escape

In the query field, clears the query and returns to suggestions. In the settings popover or the Add source menu, closes it and returns focus to its trigger.

Space

Toggles the Rerank, Hybrid search and Reranking switches and activates buttons.

• The card is a labelled section. Each pipeline is a progressbar with the overall percent and a value text that names the step, such as Embed, step 3 of 4, or Failed at embed.
• Status never relies on color: indexed rows show a check with a label, failed rows an alert icon with the reason in text, and working rows a named step and a percent.
• Each result is a button with aria-expanded and aria-controls; its similarity or reranker score and any rank change (for example, Up 2 after reranking, was 3) are in its accessible name.
• Matching words are wrapped in mark elements, so assistive technology can report highlighted text.
• Ingestion milestones, failures, search results, reranking and simulated actions are announced through one polite status region; simulated uploads say that files stay in the browser.
• Icon-only controls (settings, clear query) have specific names. The settings toggles are full labels, so the text toggles the switch too.
• Pointer focus draws no ring; keyboard focus shows Arc's shared focus ring.

• Turning on Rerank shows a short reranking state, then moves each result to its new rank with a critically damped layout animation (no bounce) while score bars and numbers roll to the reranker's values and a rank change fades in.
• Opening a result springs its height open: the previous chunk slides in above, the excerpt grows from two lines to the full chunk, and the next chunk slides in below, all on the smooth spring.
• Pipeline segments fill from a single animation-frame clock while counts climb; the meta line crossfades when a step changes. Active labels use a finite text shimmer that settles when the step ends.
• New sources open in at the top of the list on a height spring; removed ones close their gap. Views crossfade without moving.
• Reduced motion places results, heights and values immediately, stops the spinner and the skeleton pulse, and keeps every label and announcement.

• The card is container aware: from about 720px down to 320px it keeps one column, with no horizontal scroll.
• Below 560px the pipeline moves under each source name at full width, step labels drop, the page or path beside a result shortens, and the chunk-in-context rail moves inline.
• Below 440px the gutters tighten, the header actions wrap under the title, meta lines wrap instead of truncating, and the indexing summary beside the view switch hides.
• Long source names truncate with the full name in a title; failure reasons always wrap so they stay readable.

• The simulation advances all sources from one requestAnimationFrame loop that only runs while something is queued or working.
• Scores and counts roll through motion values without React re-renders; excerpt heights are measured once with a ResizeObserver and only re-measured when their width changes.
• The local retriever scores a few dozen chunks with plain string matching; real deployments pass onQuery and skip it.

• Pass sources from your ingestion job and update status and progress as events arrive; the card only renders them when sources is set. Leave sources unset for a self-running demo.
• Return more candidates than top k from onQuery (for example 20) with both similarity and relevance. That is what lets the reranker pull a chunk up from below the cut, which is the point of the comparison.
• Scores are expected between 0 and 1. Map cosine similarity or a reranker logit into that range before returning it.
• Give each result before and after neighbours (or index plus a chunks prop) so opening a result can show it in context.
• location is free text: a page such as "p. 3" for documents, a heading for Notion, or a path for a crawled site.
• The embedding model labels are generic placeholders. Replace them with the models your stack actually uses.

The full library index for assistants is at /llms.txt.