Review mining agent

This page explains how a review mining agent reads ecommerce, App Store, G2, or review dashboard sources under a read-only grant, extracts customer feedback from authorized and public reviews, clusters it into themes, and produces positioning insights. After reading it, you will understand which modules this scenario needs, how the permission boundary narrows, and why review verbatims and cluster results belong in your analytics store rather than Memory.

Use case

Marketing and product teams need to understand purchase motivations, objections, feature feedback, and competitor comparisons from customer reviews. Reviews are scattered across platforms, high-volume, and always growing — reading them manually cannot produce reviewable theme clusters, and "users are all saying" conclusions without sources cannot support positioning decisions.

Typical triggers:

• After a release or product launch, review themes across platforms must be summarized within a week.
• Before a positioning retrospective, purchase motivations and objections need evidence in users' own words.
• A competitor's rating shifts, and feature feedback must be compared across both products' reviews.

Engineering challenges

• High volume and high noise: thousands of cross-platform reviews mix fake reviews, off-topic complaints, and duplicates, so clustering easily produces conclusions that "look reasonable" without a single supporting verbatim.
• Insights must be traceable: a conclusion like "users all say it's too expensive" cannot be re-verified without review source URLs, and cannot defend itself when the positioning decision is challenged — an unsourced insight is no insight.
• The theme structure must stay comparable across cycles: change the classification rules and two cycles' theme distributions can no longer be compared; rules must be versioned, and cluster results must be archived together with the rule version for trend judgments to mean anything.

Module composition

Module	Role	Notes
GenAuth	Core	Read-only delegation, revocation, and the audit chain for authorized review dashboards; the dashboard account's reply, report, and merchant-info powers stay outside the delegation.
Web Agent	Core	Controlled sessions sign in to review dashboards or visit public review pages, extracting ratings, verbatim excerpts, and source links page by page; public review pages can be located with WebSearch.
GUMem	Not used	Review verbatims, cluster versions, and trend data are analytics assets — archive them to your analytics store. Classification rules and competitor mappings are versioned policy — keep them in your policy store and inject them per version. Neither is Memory.

Permission and delegation boundaries

The Agent holds no inherent permissions. The effective authority for each task is the intersection of three sets: what the user actually holds ∩ what was explicitly delegated for this task ∩ what the enterprise has approved. Applied to this scenario:

• The delegated scope covers only "read authorized review dashboards and public review pages for the selected products" — no replying to reviews, reporting content, or editing merchant info.
• Delegation credentials are short-lived; minute-level validity is recommended for a single mining run.
• The user or an administrator can revoke the grant at any time; new review-extraction requests fail immediately after revocation.
• Out-of-scope attempts (for example, another product's review dashboard outside the delegation) are rejected and recorded — the audit chain covers all attempts, not just successful actions.

Note: this sample requests doAnything product permissions. Your application and downstream services must configure and enforce business limits such as platform lists, product ranges, action whitelists. Product scopes and prompt rules do not enforce those detailed limits; this sample does not configure them. See Delegate token and attenuation.

Workflow

1. The user selects a product, platform, and time range and confirms the interactive authorization in Qoni Console.

2. GenAuth issues a least-privilege, read-only delegation credential for this task.

3. Your app loads the current version of the classification rules, theme taxonomy, and competitor mappings from the policy store and injects them into the task.

4. Web Agent signs in to review dashboards or opens public review pages and extracts ratings, verbatim excerpts, and source links page by page.

   Checkpoint: When a review platform shows a login wall, CAPTCHA, or risk-control page, the Web Agent should escalate to a human instead of silently bypassing it.

5. The Agent clusters reviews into themes — purchase motivations, objections, feature feedback, and competitor comparisons — each insight with a representative verbatim and source link.

6. Your app validates the output contract — insights missing a review source URL are dropped — and archives cluster results with the rule version to your analytics store.

7. The Agent returns theme clusters, representative excerpts, messaging suggestions, and a source list, with the taxonomy version and an audit id.

   Checkpoint: Every sentiment or theme conclusion should trace back to concrete review source links; conclusions without excerpts should not enter the deliverable.

Example code

This example uses @qoniai/qoni 0.9.0, published on npm. The download includes the same SDK version, installed with npm ci. The demo groups public Mozilla support questions by usability, privacy, and compatibility. The SDK reads real public pages; business inputs in scenarios.ts are labeled public-demo.

This demo does not read or write GUMem; its task uses the supplied page list and explicit business inputs. DoAnything opens the supplied pages and produces the scenario output. Supply your own JSON with --input; use --interactive when the user must consent to delegation. For site sign-in and user responses, see Qoni SDK.

Download the complete runnable examples, or run from the documentation repository:

cd examples/qoni
npm ci
npm run case -- review-mining-agent
# Supply your own inputs
npm run case -- review-mining-agent --input /path/to/input.json

Set server-side QONI_ACCESS_KEY and QONI_SECRET_KEY. QONI_USER_ID can identify your application's current GenAuth user; the local demo otherwise selects a user from the bound pool. Demonstration Memory writes use an isolated user rather than changing a business user's preferences.

This scenario's executable entry point:

import { cliOptions } from '../runtime.js'
import { inputFile, runScenario } from '../run-case.js'

// Extract themes from public feedback and retain inspectable quotations.
const report = await runScenario('review-mining-agent', cliOptions(), inputFile())
// report.items: themes, insights, quotations, and links to the original feedback.
// Validated fields: theme, insight, quote, reviewUrl.
console.log(JSON.stringify(report, null, 2))

The entry point loads the definition below by scenario ID. The code is included directly from scenarios.ts, with comments shown in the page language: the task, output fields, source field, Web Search queries (if any), whether Memory is used, and the demonstration input. The pipeline appends the input data, search sources, recalled Memory, and shared safety constraints to the task to build the final prompt; see the pipeline below for the full assembly.

// Extract themes from public feedback and retain inspectable quotations.
browser('review-mining-agent',
  'Read this public support discussion as a demonstration of feedback classification. Return [{theme,insight,quote,reviewUrl}]. Use actual visible user-feedback text, retain its source, and exclude names or other personal data. Do not reply or edit the discussion.',
  ['theme','insight','quote','reviewUrl'], 'reviewUrl', sample(['https://support.mozilla.org/en-US/questions/'], { taxonomy:['usability','privacy','compatibility'] })),

The sample implements runScenario(), browser(), research(), and sample() as application functions. The pipeline below makes the actual SDK calls: delegation and introspection → required Memory/search → browser task or monitor → validation and saving. fields and sourceField define the application's output checks. The complete application helpers are in the package's runtime.ts.

Inspect the actual SDK pipeline

// browser() uses DoAnything; research() searches first; sample() labels public-demo inputs.
const firefox = 'https://www.mozilla.org/en-US/firefox/new/'
const manifesto = 'https://www.mozilla.org/en-US/about/manifesto/'
const privacy = 'https://www.mozilla.org/en-US/privacy/firefox/'
const support = 'https://support.mozilla.org/en-US/kb/get-started-firefox-overview-main-features'
// These copy rules become prompt context; server permissions and business checks remain separate.
const policy = {
  version: 'demo-2026-09',
  approvedClaims: ['Describe only features supported by the cited page.'],
  forbiddenClaims: ['guaranteed security', '100% private', 'unverified pricing or performance'],
  voice: 'concise and warm',
}
const sample = (pages: string[], business: JsonObject = {}): JsonObject => ({
  dataset: 'public-demo', pages, policy, business,
  notice: 'Business records are synthetic demonstration inputs. Referenced websites and SDK execution are real.',
})
// fields lists required output keys; sourceField identifies URL checks; memory enables GUMem calls.
const browser = (id: string, task: string, fields: string[], sourceField: string | undefined, input: JsonObject, memory = false): Scenario => ({
  id, products: ['doAnything'], task, fields, sourceField, input, memory,
})
const research = (id: string, task: string, fields: string[], sourceField: string, queries: string[], input: JsonObject, memory = false): Scenario => ({
  id, products: ['webSearch', 'doAnything'], task, fields, sourceField, queries, input, memory,
})

import { QoniScopes, type JsonObject, type RunResult } from '@qoniai/qoni'
import { readFileSync } from 'node:fs'
import { getScenario, type Scenario } from './scenarios.js'
import { appendTrace, checkInputCoverage, cleanupDemoUser, createContext, delegate, handleInteraction, inputEntryCount, isolateDemoUser, object, readWithRetry, renderScreenshot,
  save, saveArtifacts, searchHits, settled, settleRun, validateItems, withCleanup, type Context, type Options } from './runtime.js'

export async function runScenario(id: string, options: Options = {}, input?: JsonObject) {
  // Load the task definition by ID; --input replaces its business inputs.
  const scenario = getScenario(id)
  const data = input ?? scenario.input
  if (scenario.memory && data.dataset !== 'public-demo' && options.mode !== 'interactive' && !options.userId && !process.env.QONI_USER_ID) {
    throw new Error('Business Memory writes require the current QONI_USER_ID; do not select an arbitrary bound user')
  }
  const context = await createContext(id, { ...options,
    skipUserResolution: scenario.memory && data.dataset === 'public-demo' })
  return withCleanup(context, async register => {
    register('isolated demonstration user', () => cleanupDemoUser(context))
    // Isolate demo preferences; business Memory belongs to the identified current user.
    if (scenario.memory && data.dataset === 'public-demo') await isolateDemoUser(context)
    return await executeScenario(context, scenario, data)
  })
}

export async function executeScenario(context: Context, scenario: Scenario, input: JsonObject) {
  const pages = input.pages
  if (!Array.isArray(pages) || !pages.length || pages.some(page => typeof page !== 'string' || !/^https:\/\//.test(page))) {
    throw new Error('Input pages must be an array of HTTPS URLs')
  }
  if (input.requiresLogin === true && context.mode !== 'interactive') {
    throw new Error('Targets that require sign-in need --interactive and user-controlled login')
  }
  const memoryScopes = scenario.memory
    ? [QoniScopes.GUMEM_MEMORY_READ, QoniScopes.GUMEM_MEMORY_WRITE, QoniScopes.GUMEM_MESSAGE_WRITE] : []
  // delegate() is an application helper around the SDK delegation methods.
  const grant = await delegate(context, scenario.id, scenario.products, memoryScopes)
  // Read the effective scopes; readWithRetry() retries only retryable read failures.
  const { data: tokenInfo } = await readWithRetry(context, 'delegation introspection',
    () => context.qoni.genauth.introspectDelegationToken({ token: grant.token }))
  const info = object(tokenInfo)
  if (info.active !== true) throw new Error('The delegation token is not active')
  const audit = { grantId: grant.grantId, auditId: grant.auditId, scopes: info.scope }
  let memory: unknown
  if (scenario.memory) {
    // A Session associates this conversation with the user; the app chooses sessionId.
    const sessionId = `${scenario.id}-${Date.now()}`
    await context.qoni.gumem.createSession({
      token: grant.token, userId: context.userId, sessionId, title: scenario.id,
    })
    const preferences = object(input.business ?? {}).confirmedPreferences
    if (Array.isArray(preferences) && preferences.length) {
      // Store confirmed preferences only; sync: true requests synchronous processing.
      await context.qoni.gumem.addMessages({ token: grant.token, userId: context.userId, sessionId, sync: true,
        messages: [{ role: 'user', content: `Confirmed demonstration preferences: ${preferences.join('; ')}` }] })
    }
    // Recall relevant preferences for the later task prompt.
    memory = (await readWithRetry(context, 'GUMem recall', () => context.qoni.gumem.recall({ token: grant.token, sessionId,
      query: 'Confirmed preferences relevant to this task', details: true }))).data
    save(context, 'memory.json', { sessionId, context: memory })
    if (Array.isArray(preferences) && preferences.length && !preferences.every(value => JSON.stringify(memory).includes(String(value)))) {
      throw new Error('Recall did not include the confirmed preferences just written by this demo')
    }
  }

  if (scenario.products.includes('track')) return runMonitor(context, scenario, input, grant.token, audit)

  let hits: ReturnType<typeof searchHits> = []
  if (scenario.queries) {
    // Web Search returns results[]; DoAnything receives these sources to inspect.
    const search = await context.qoni.webSearch.run({ token: grant.token, prompt: scenario.queries, maxResultsPerQuery: 3 })
    save(context, 'search-ref.json', { runId: search.id, audit })
    await withCleanup(context, async register => {
      register('Web Search run', () => search.cancel('Documentation demonstration cleanup'))
      const result = await settleRun(context, search)
      save(context, 'search-result.json', result)
      settled(result)
      hits = searchHits(result.output)
    })
  }

  // One-per-entry scenarios request exactly one item per input entry; others at most two.
  const requiredItems = inputEntryCount(scenario.id, input)
  // Assemble the task, inputs, search sources, and Memory as application-defined context.
  const prompt = [scenario.task, `Task inputs: ${JSON.stringify(input)}`,
    `Search sources: ${JSON.stringify(hits)}`, `Confirmed memory: ${JSON.stringify(memory ?? null)}`,
    `Actual collection time: ${new Date().toISOString()}`,
    requiredItems === undefined
      ? 'Inspect the supplied sources. Return at most two items in the requested JSON array, without prose or Markdown.'
      : `Inspect the supplied sources. Return exactly ${requiredItems} item${requiredItems === 1 ? '' : 's'} in the requested JSON array, one per input entry, without prose or Markdown.`,
    'Keep synthetic demonstration data identified as synthetic. Do not send messages, publish, pay, edit accounts or submit forms.',
    input.requiresLogin === true ? 'Request user sign-in through an interaction when required; never enter credentials yourself.' : 'Public demonstration sources only; do not sign in.',
  ].join('\n\n')
  // Start the Agent with this grant; capture receives delivered screenshots, not every step.
  const run = await context.qoni.doAnything.run({ token: grant.token, prompt, capture: { screenshots: true } })
  save(context, 'run-ref.json', { runId: run.id, session: run.sessionRef, audit })
  let result: RunResult
  const trace = (event: { type: string; data: unknown }) => {
    context.eventCounts[event.type] = (context.eventCounts[event.type] ?? 0) + 1
    if (['progress','message','done'].includes(event.type)) appendTrace(context, event)
    if (event.type === 'browserLiveUrlChanged') {
      const liveUrl = object(event.data).liveUrl
      if (typeof liveUrl === 'string') context.browserUrl = liveUrl
    }
  }
  return withCleanup(context, async register => {
    register('DoAnything run', () => run.cancel('Documentation demonstration cleanup'))
    if (context.delivery === 'events') {
      // --events streams updates; wrap interaction data in an SDK handle for user handling.
      for await (const event of run.events({ signal: AbortSignal.any([context.abort.signal, AbortSignal.timeout(context.timeoutMs)]) })) {
        trace(event)
        if (event.type === 'screenshot') renderScreenshot(context, event.image)
        if (event.type === 'interaction') await handleInteraction(context, run.interactionHandle(event.data))
      }
      result = await settleRun(context, run)
    } else {
      // Callback mode receives this run's events inside wait; helpers save images and ask the user.
      result = await settleRun(context, run, { onEvent: trace,
        onScreenshot: (image, index) => renderScreenshot(context, image, index),
        onInteraction: interaction => handleInteraction(context, interaction) })
    }
    save(context, 'result.json', result)
    settled(result)
    await saveArtifacts(context, result)
    if (scenario.id === 'landing-page-audit-agent' && context.screenshots === 0) throw new Error('The landing-page audit did not deliver the requested screenshot')
    // The app checks required fields and source URL formats; a reviewer still checks facts.
    const items = validateItems(result.output, scenario.fields, scenario.sourceField)
    // Scenarios that require one item per input entry are checked against the input.
    checkInputCoverage(scenario.id, items, input)
    const report = { scenario: scenario.id, dataset: input.dataset, passed: true, runId: run.id,
      status: result.status, items, audit, screenshots: context.screenshots, interactions: context.interactions,
      events: context.eventCounts, artifactIds: result.artifacts.map(artifact => artifact.id) }
    save(context, 'report.json', report)
    return report
  })
}

async function runMonitor(context: Context, scenario: Scenario, input: JsonObject, token: string, audit: JsonObject) {
  // Track creates a monitor with targets, extraction fields, and hourly scheduling.
  const monitor = await context.qoni.track.create({ token, prompt: scenario.task,
    targetUrls: input.pages, extractionSchema: { heading: 'string', source_url: 'string' },
    tickInstructions: `Open the target URLs and read the actual visible heading. Return a JSON object with heading and source_url. ${scenario.task}`,
    triggerDsl: { on: 'change' }, schedule: { kind: 'interval', intervalSeconds: 3600 } })
  save(context, 'monitor-ref.json', { id: monitor.id, audit })
  return withCleanup(context, async register => {
    register('Track monitor', () => monitor.delete())
    const definition = await monitor.get()
    save(context, 'monitor-definition.json', definition)
    if (object(definition.schedule).intervalSeconds !== 3600) throw new Error('Track did not persist the requested schedule interval')
    // Run one tick and inspect its extraction by runId; completed alone does not prove success.
    const tick = await monitor.runNow()
    save(context, 'tick.json', tick)
    if (tick.state !== 'completed') throw new Error(`Track execution failed: ${tick.state} / ${tick.error ?? ''}`)
    const runId = tick.runId
    if (typeof runId !== 'string') throw new Error('Track tick did not return a runId')
    const detail = await monitor.run(runId)
    save(context, 'tick-detail.json', detail)
    if (detail.state !== 'completed' || !detail.extracted || !Object.keys(object(detail.extracted)).length) {
      throw new Error('Track did not extract page data')
    }
    const extracted = object(detail.extracted)
    if (typeof extracted.heading !== 'string' || !extracted.heading.trim() ||
      typeof extracted.source_url !== 'string' || !/^https:\/\//.test(extracted.source_url)) {
      throw new Error('Track extraction is missing a heading or source URL')
    }
    const normalizeUrl = (value: string) => { const url = new URL(value); url.hash = ''; return url.href.replace(/\/$/, '') }
    if (!(input.pages as string[]).some(url => normalizeUrl(url) === normalizeUrl(String(extracted.source_url)))) {
      throw new Error('The Track source URL is not a configured target')
    }
    // Check persisted pause/resume state; withCleanup() deletes the monitor on exit.
    await monitor.pause()
    if ((await monitor.get()).status !== 'paused') throw new Error('Track did not persist the paused state')
    await monitor.resume()
    if ((await monitor.get()).status !== 'active') throw new Error('Track did not persist the active state')
    const report = { scenario: scenario.id, dataset: input.dataset, passed: true, monitorId: monitor.id,
      runId, state: detail.state, outcome: detail.outcome, extracted: detail.extracted, audit }
    save(context, 'report.json', report)
    return report
  })
}

export function inputFile(): JsonObject | undefined {
  const index = process.argv.indexOf('--input')
  return index >= 0 ? object(JSON.parse(readFileSync(process.argv[index + 1], 'utf8'))) : undefined
}

Results are written to output/review-mining-agent/report.json. report.items contains themes, insights, quotations, and links to the original feedback, with fields theme, insight, quote, reviewUrl. audit links the grant ID, audit ID, and effective scopes; http.json records redacted request statuses. The application parses DoAnything output and checks required fields and source URL formats. A business reviewer still assesses the content against the original sources.

Data and memory boundaries

This scenario touches four kinds of data; none of them needs GUMem:

• Versioned rules: classification rules, the theme taxonomy, competitor mappings — managed by version in your policy store, with every cycle's clusters citing the rule version.
• Business state: review verbatims, cluster results, trend data — archived by rule version to your analytics store for cross-cycle comparison and review.
• Audit records: the delegation and behavior chain formed by grantId and auditId — maintained by GenAuth.
• User Memory (optional): this scenario neither recalls nor writes back by default; review content is the reviewers' public expression and should not accumulate as anyone's Memory.

Failure handling

Situation	Recommended handling
Review dashboard login state expires	Suspend the task, notify the user to sign in again, and resume from the checkpoint.
Review page structure changes break extraction	Treat it as a failure and replay the session recording; never emit insights without evidence.
Request for a product dashboard outside the delegated scope	Reject and record it; the attempted access remains visible in the audit chain.
An insight lacks a review source URL	App-side validation drops the entry and the report notes how many were dropped.

Production notes

Do not expose private user information. Public review quotes should follow platform rules and use minimal necessary excerpts. The Agent only processes publicly visible or authorized review content; reviewer personal data is used for insight analysis only, never for outreach or profiling, and extraction frequency against public review pages should be capped in line with platform terms of service.

Next steps

• Read the Quickstart to run the shortest path for Agent identity and delegation.
• Read Authorization and browser sandbox for the security boundaries of controlled sessions.
• Continue with the Influencer vetting agent for an adjacent scenario.