Platform Mode

GuideKit Platform Mode wires the optional intelligence, knowledge, and plugins packages into the v2 pipeline — semantic page enrichment, RAG retrieval, plugin hooks, and hallucination validation.

For full-site assistance, Platform Mode now composes three layers:

Site knowledge: server-backed search across indexed website documents via /api/guidekit/site-search
Live page model: current route sections, forms, overlays, and interactive elements from the DOM scanner
Autonomy policy: guided action rules for navigation, safe clicks, and confirmation-required actions

Minimal vs Platform Mode

Mode	Packages	Use case
Minimal	`@guidekit/core`, `@guidekit/react`, `@guidekit/server`	Chat widget with DOM scan + tools
Platform	+ `@guidekit/intelligence`, `@guidekit/knowledge`, `@guidekit/plugins`	Production agent with semantic model, RAG, extensions

Enable Platform Mode


import { GuideKitProvider } from '@guidekit/react';
import { definePlugin } from '@guidekit/plugins';
 
const docs = [
  { id: '1', title: 'Billing', content: 'Manage billing in Settings > Billing.', tags: ['billing'] },
];
 
const myPlugin = definePlugin({
  name: 'my-plugin',
  version: '1.0.0',
  setup: () => {},
});
 
<GuideKitProvider
  tokenEndpoint="/api/guidekit/token"
  proxy={{ llm: '/api/guidekit/llm', health: '/api/guidekit/health' }}
  llm={{ provider: 'gemini', model: 'gemini-2.5-flash' }}
  intelligence={true}
  knowledge={{ documents: docs, engine: 'bm25', topK: 5 }}
  siteKnowledge={{ endpoint: '/api/guidekit/site-search', topK: 5 }}
  plugins={[myPlugin]}
  hallucinationGuard
  options={{
    autonomy: {
      level: 'guided',
      allowNavigation: true,
      allowSafeClicks: true,
      requireConfirmationFor: ['submit', 'purchase', 'destructive', 'auth'],
    },
  }}
>
  {children}
</GuideKitProvider>

Pipeline stages

scan — DOM page model
enrich — SemanticScanner → SemanticPageModel
retrieve — BM25/TF-IDF knowledge search plus optional server site search
context — System prompt with semantic sections + RAG
cognize — Optional cognitive planning when cognitive={true} (Tier C experimental)
llm — Proxy-backed streaming (includes tool execution)
validate — Hallucination guard confidence (advisory only in 1.0)
render — Store + events

Plugin hooks

Plugins registered via @guidekit/plugins can implement lifecycle hooks that the v2 pipeline invokes:

Hook	When it runs
`beforeScan` / `afterScan`	Around DOM page model capture
`beforeEnrich` / `afterEnrich`	Around semantic enrichment (Platform Mode)
`beforeRetrieve` / `afterRetrieve`	Around knowledge retrieval
`beforeLLMCall` / `afterLLMCall`	Before and after each LLM round
`onToolCall`	When a tool is about to execute
`onError`	On pipeline or tool errors

Use getExtraContextSections() on a plugin to inject additional system-prompt sections during the context stage.

Knowledge limits (1.0)

Documents are held in memory on the client for the session — there is no IndexedDB persistence in 1.0.
Use addKnowledgeDocument / removeKnowledgeDocument on the core instance (or React context) for runtime updates.
Large corpora should live server-side; pass a curated subset via knowledge.documents at init time.
Use siteKnowledge.endpoint for full-site context. The server route validates the GuideKit JWT, site:read permission, and allowed origins before returning attributed website chunks.

Agent Runtime v1

The built-in agent can search indexed website content with searchSite, navigate same-origin routes, rescan after route changes, and act on current-page elements. Interactive elements are classified as safe, submit, purchase, destructive, auth, or unknown; guided autonomy allows safe clicks and requires confirmation for risky classes by default.

Reference integration

See apps/example-nextjs for the canonical Platform Mode demo with zero client API keys.