Dashboard Overview
The screen an app opens on: shell, header, KPI row, chart and activity feed — numbers first, shape second, changes last.
Why the answer said what it said: what is connected, how stale it is, what the query was allowed to see, what fitted in the window and what fell out — plus the failed-search state a debugging tool is most needed in.
app/retrieval-console-page.tsxnpx hoverlab add retrieval-console-page
Or over MCP, from your editor's agent — no account needed.
Free to read, copy and install, for personal and non-commercial projects. Shipping it in client work or a paid product needs Pro ($79 once). The source lands in your repo and stops being ours — no attribution, nothing to upgrade.
What is connected, how current it is, what the query could see, and what actually made it into the answer.
Answers can only be as current as the source they come from. The oldest connected source was last read 6 weeks ago, so anything changed since then may not be reflected yet.
Website · 4 domains
Up to date · last read 11 minutes ago
486 docs
Notion workspace
Partly indexed · last read 2 hours ago
812 of 1,340 documents searchable — the remaining 528 cannot be cited yet.
1,340 docs
Zendesk · resolved only
Behind · last read 9 days ago
22,190 docs
Google Drive folder
Not syncing · last read 6 weeks ago
The connected account lost access to the folder on 3 August. The 74 documents already indexed are still being answered from, at the age above.
74 docs
Relevance ranking says which document matched. It says nothing about whether the document is still true, and that is the question a reader has once the answer surprises them.
Citations tell you what an answer used. Scope tells you what it could have used, which is the question behind "why did it not know that".
4 chunks for “why did northwind not get the volume price”, ranked by similarity
Accounts above 1,200 units per order qualify for the tier-two rate of $8.40, held for the duration of the contract year. The break is applied per order, not per quarter, so a single large order beats two smaller ones of the same total volume.
Customer asked why their November order did not get the tier-two price. Their October and November orders were 700 units each — combined they clear the break, individually they do not. Confirmed with finance that this is working as intended.
Aggregated order volume by account for the trailing four quarters, filtered to accounts with more than one order in the period.
Standard lead time for the winter blend is 21 days from purchase order. Expedited shipping is available at a 15% surcharge and reduces this to 9 days.
27.0k of 32.0k tokens · 84%
The model did not see any of this. If an answer looks like it ignored a document, start here.
Database table: Q3 revenue cohortswarehouse.arr_monthly
Supports: The 6.1% churn figure for accounts under 20 seats
2,463 accounts · lost ARR $1.24M · seat_band < 20
Read
Database table: Activation funnel, Jul–Sepwarehouse.events
Supports: The claim that a second integration predicts retention
Accounts with ≥2 connected sources churned at 1.6%; with 1 source, 6.4%.
Read
Document: Pricing handbookdocs/pricing-handbook.md
Supports: That the volume break applies per order rather than per quarter
The break is applied per order, not per quarter, so a single large order beats two smaller ones of the same total volume.
Read
This document has changed twice since it was indexed.
Web page: Supplier lead times (opens in a new tab)supplier.example.com
Supports: The 21-day reorder window
Read
Every figure above traces to one of these. Nothing was inferred without a source.
Rather than guess at a number that has to be exact, here is where I looked and what came closest.
what is our EU VAT registration number
Tax registrations usually live in the finance drive, which is not connected.
The real page, rendered in your current theme — every section below is a live block, not a screenshot.
Only want one section? Open it and copy that block instead — browse Retrieval & Context.
/**
* Why the answer said what it said — the RAG side of an assistant, as an
* operable screen.
*
* The question this page exists to answer is not "is retrieval working".
* It is "is retrieval working *now*, on the documents I think it has, and
* did the thing it just told me actually come from one of them". Those are
* three different questions and most consoles answer only the first.
*
* So the order is: what is connected, how stale it is, what it was allowed
* to look at, what it found, what fitted, and what it cited.
*
* index status a green tick means the connector works, which is not
* the same as the answers being current
* freshness so the previous sentence has somewhere to be answered
* scope what the query was permitted to see. A retrieval bug
* and a permissions boundary look identical from the
* outside, and only this list tells them apart
* chunks the passages themselves, ranked
* budget what fitted in the window and what fell out of it —
* the most common cause of "it ignored the document I
* gave it", and invisible without this
* citations the published footnotes, which is the only part of
* this page an end user ever sees
*
* The empty state is last and it is not an afterthought. A retrieval
* console that cannot show you a failed search is a console you can only
* use when you do not need it: <RetrievalEmptyState> says plainly that
* nothing was found, where it looked, and what the near-misses were, which
* is the difference between debugging and guessing.
*/
import * as React from 'react'
import { RetrievalIndexStatus } from '@/components/retrieval-index-status'
import { RetrievalFreshnessList } from '@/components/retrieval-freshness-list'
import { ContextScopeList } from '@/components/context-scope-list'
import { ContextChunkCards } from '@/components/context-chunk-cards'
import { ContextWindowBudget } from '@/components/context-window-budget'
import { SourceCitationList } from '@/components/source-citation-list'
import { RetrievalEmptyState } from '@/components/retrieval-empty-state'
export default function RetrievalConsolePage() {
return (
<main className="min-h-screen bg-background text-foreground">
<section className="mx-auto w-full max-w-5xl px-6 pb-2 pt-12">
<h1 className="text-2xl font-bold tracking-tight">Retrieval</h1>
<p className="mt-1 text-sm text-muted-foreground">
What is connected, how current it is, what the query could see, and
what actually made it into the answer.
</p>
</section>
<RetrievalIndexStatus />
<RetrievalFreshnessList />
<ContextScopeList />
<ContextChunkCards />
<ContextWindowBudget />
<SourceCitationList />
{/* The state a debugging tool is most needed in. */}
<RetrievalEmptyState />
</main>
)
}
components/ — each block page has its own copy button.app/retrieval-console-page.tsx. The imports already point at @/components/…, so they resolve with no edits.The component, its props, the design tokens it expects and the command that installs it — as one prompt. Paste it into Claude, Cursor, v0 or ChatGPT and what they build around it will match the rest of the catalog instead of inventing its own system.
The screen an app opens on: shell, header, KPI row, chart and activity feed — numbers first, shape second, changes last.
An agent working inside a real app: a transcript down the middle — reasoning, answer, then the one card that can act — with what it may read and what it noticed demoted to a rail.
A kanban board inside the app shell, with tabs that admit the board is one view among several and a header whose primary action is creating work.
The canonical CRUD list — header, toolbar, sortable table and pagination framed as one panel instead of three stacked cards.
A wizard for the session someone finishes and a checklist for the one they do not, with the only genuinely blocking step in front of both.
Built around the state a search page is in most of the time — empty — so recent queries lead and the facets sit in a sidebar that survives growing to twelve.
One run opened up in the order you would debug it — intent, actions, the retries that are invisible in both, then cost.
The queue, the same queue sorted by how long things have waited, and the policy that decided both — with the policy last so it is not scrolled past forever.
The blank thread answered with real starter prompts, the model picker with the three numbers the choice turns on, and the four surfaces a session grows into — attachments, branches, canvas, threads.
The three states an agent screen usually omits: the wait before the first token, the trace after something broke with retries costed in the open, and the diff it wants a person to accept.
Five ways to reach a model without the cursor leaving the sentence — slash menu, inline completion, selection toolbar, action menu, property inspector — and the two surfaces that say where the answer came from.
A table after real people have used it — columns reordered from the keyboard, a selection bar that asks page-or-all, subtotals that survive collapsing, cells edited in place, and the empty state its own filters produced.
The saved view and the comparison window first, because every figure below depends on both — then six charts, none of which load a charting library.
Thresholds with how often each has actually fired, what that produced, where it went — and the digest settings that exist because of the answer.
The flow between an empty product and a useful one, in the order it goes wrong: sample data offered with equal weight, a file, a URL, the column mapping everything stalls on, and an upload that survives the connection dropping.
A profile form where the role options state what they actually change, the avatar crop works from the keyboard, and a navigation guard names the fields it is about to throw away.
The furniture that is on every screen and belongs to none — navbar, mobile drawer, command palette, shortcut sheet, header search, scope switcher, and the two shapes a detail view takes without losing the list.
Backwards from now and forwards from it on one screen — a day-grouped timeline, a month grid where a busy Tuesday never changes the row height, and the state both are in on day one.
What the product looks like while onboarding happens: a tour step anchored to the control it explains, the skeletons of the slowest load it will ever do, and the checklist that outlives both.