Retrieval inspector

Pro

The passages behind an answer, in the order the retriever ranked them. Each says whether the answer actually cites it — retrieved-but-unused is shown, not hidden — with its score (or that it wasn't scored), its source and how fresh it is. Expand a passage for the excerpt and why it ranked there.

Markdown

Preview

Install

With the Candor CLI

First time? Get started sets up the registry and theme. Or with shadcn directly: npx shadcn@latest add @candor/retrieval-inspector

Brings in: collapsible, freshness-stamp, know, knowledge-sources, locale, utils · npm: lucide-react@^1.67.0

Usage

TSX · The example above

"use client"

import type { KnowSource, RetrievedPassage } from "@/lib/know"
import { RetrievalInspector } from "@/components/retrieval-inspector"

const ago = (h: number) => new Date(Date.now() - h * 3_600_000).toISOString()

const sources: KnowSource[] = [
  { id: "k1", door: "config", label: "Staging: code and schema", readAt: ago(5), everyHours: 24, status: "succeeded" },
  { id: "k2", door: "data", label: "customers, sampled", readAt: ago(5), everyHours: 168, status: "succeeded" },
  { id: "k3", door: "file", label: "Data model notes v3.pdf", addedBy: "Priya", readAt: ago(24 * 12), status: "succeeded" },
  { id: "k4", door: "answer", label: "Answer from Priya", addedBy: "Priya", readAt: ago(24 * 9), status: "succeeded" },
]

const passages: RetrievedPassage[] = [
  { id: "p1", sourceId: "k3", title: "§3.2 Data regions", excerpt: "Customers are routed to a data region by region (us, eu, apac). Partner tier is no longer used for routing.", score: 0.91, cited: true, why: "Matches “data region” and “partner tier”", readAt: ago(24 * 12) },
  { id: "p2", sourceId: "k1", title: "Job: assign_region", excerpt: "assign_region reads customers.region; the partner-tier fallback sits behind a flag that is off.", score: 0.84, cited: true, why: "The job named in the answer", readAt: ago(5) },
  { id: "p3", sourceId: "k4", title: "Answer from Priya", excerpt: "Partner tier is being replaced by region. Keep it on the two partner dashboards until Q1.", score: null, cited: true, why: "Pinned: a person's answer to this exact question", readAt: ago(24 * 9) },
  { id: "p4", sourceId: "k2", title: "Fill rates on customers", excerpt: "region is filled on almost every sampled row; partner_tier_old has no recent writes.", score: 0.71, cited: false, why: "Relevant, but the answer didn't need it", readAt: ago(5) },
]

export default function Example() {
  return <RetrievalInspector className="max-w-2xl" passages={passages} sources={sources} />
}

<RetrievalInspector>

Prop	Type	Default	Description
passagesrequired	`RetrievedPassage[]`	—	The retrieved passages, in rank order, each marked cited or not.
sourcesrequired	`KnowSource[]`	—	The sources the passages come from, looked up by id for name and freshness.
verb	`string \| undefined`	—	The word before each passage's date, for what `readAt` records: "Reviewed", "Published", "Synced". Defaults to FreshnessStamp's "Read".
className	`string \| undefined`	—	Extra classes for the outer element.

Strings

The fixed words, in English by default. Change any of them under retrievalInspector in <LocaleProvider strings={{ … }}>; the table is exported as retrievalInspectorStrings.

Key	English
label	Retrieved passages
heading	What was retrieved
passages	{ one: "{count} passage", other: "{count} passages" }
cited	{count} cited
citedOnly	Cited only
citedBadge	Cited
notUsed	Not used
notScored	Not scored
why	Why here: {why}