Baiyuan RAG Knowledge Platform Whitepaper

Chapter 9 — Integration with Baiyuan GEO

GEO makes brands appear in AI answers. RAG ensures the AI sees the right facts. They are two sides of one coin.

9.1 Why Deep Integration

GEO Platform (see sister whitepaper at https://github.com/baiyuan-tech/geo-whitepaper) handles seven-dimension AI citation scoring, AXP shadow docs, hallucination remediation loop. RAG handles L1 Wiki + L2 retrieval, multi-tenant knowledge. They share brand facts.

flowchart TB
    subgraph BRAND[Single Source of Brand Truth]
        ENT[brand_entities<br/>Org/Service/Person]
        FACTS[brand_facts<br/>verified claims]
    end
    subgraph RAG[RAG Platform]
        DOC[documents]
        WIKI[wiki_pages]
    end
    subgraph GEO[GEO Platform]
        AXP[AXP Shadow Doc]
        GT[Ground Truth]
        SCAN[7-dim Scan]
    end
    BRAND --> RAG
    BRAND --> GEO
    RAG -->|Wiki = GT| GT
    GEO -->|repair injection| RAG
    GEO -->|AXP embeds| WIKI

Fig 9-1: Shared model + bidirectional flow

Three data flows:

  1. brand_entities → both systems (brand bio written once)
  2. RAG Wiki → GEO Ground Truth (Wiki pages are structured facts)
  3. GEO hallucination detection → RAG injection (correct facts flow back)

9.2 Shared Brand Entity Model

CREATE TABLE brand_entities (
    id UUID PRIMARY KEY, tenant_id UUID NOT NULL,
    entity_type TEXT,  -- Organization / Service / Person / LocalBusiness
    schema_id TEXT,    -- "https://acme.example/#org"
    name TEXT, description TEXT, properties JSONB,
    sameAs TEXT[],     -- Wikipedia, Wikidata, LinkedIn
    created_at TIMESTAMPTZ, updated_at TIMESTAMPTZ
);

CREATE TABLE brand_facts (
    id UUID PRIMARY KEY, tenant_id UUID, entity_id UUID,
    claim TEXT, evidence TEXT, evidence_url TEXT,
    verified_by TEXT,  -- human / llm_nli / auto_scraped
    verified_at TIMESTAMPTZ, confidence REAL
);

brand_facts is the authoritative source. RAG Wiki and GEO GT both reference it.

9.3 Ground Truth Closed Loop

sequenceDiagram
    autonumber
    participant GEO
    participant NLI
    participant CR as ClaimReview Gen
    participant RAG
    participant Scan

    GEO->>NLI: AI claim vs brand_facts
    NLI-->>GEO: contradiction (conf=0.92)
    GEO->>CR: generate ClaimReview (schema.org)
    CR-->>GEO: JSON-LD
    GEO->>RAG: inject into wiki_pages (slug=company-overview)
    GEO->>GEO: update AXP shadow doc
    Note over Scan: 4 hours later
    Scan-->>GEO: rescan AI platforms
    Scan-->>GEO: verify correction accepted

Fig 9-2: Repair loop

Key points: NLI three-way decides only contradiction triggers repair; ClaimReview is schema.org-native; RAG is the injection point so the brand’s own CS chat also sees corrected fact.

{
  "@graph": [
    {"@type": "Organization", "@id": "https://acme.example/#org",
     "hasOfferCatalog": {"@id": "https://acme.example/#catalog"},
     "employee": [{"@id": "https://acme.example/team/alice#person"}]},
    {"@type": "Service", "@id": "https://acme.example/#service-consulting",
     "provider": {"@id": "https://acme.example/#org"}},
    {"@type": "Person", "@id": "https://acme.example/team/alice#person",
     "worksFor": {"@id": "https://acme.example/#org"}}
  ]
}

AXP shadow doc injects this JSON-LD into HTML <head>. RAG Wiki body cites the same @ids. AI crawlers (GPTBot, ClaudeBot, PerplexityBot) treat three-layer interlink as a strong knowledge-graph signal.

9.5 Hallucination Detection → RAG Auto-Repair

Scenario:

  1. Customer asks “Who is your CEO?” on chat.baiyuan.io widget
  2. RAG Wiki company-overview lacks CEO info
  3. L1 miss → L2 chunks also missing → LLM hallucinates “Bob Smith”
  4. Meanwhile GEO scans ChatGPT → also says “Bob Smith”
  5. But brand_facts says “Alice Wang”
  6. GEO triggers repair: patches RAG company-overview with CEO: Alice Wang
  7. Next customer question → Wiki hit → “Alice Wang” ✓
await rag.api.post('/api/v1/wiki/patch', {
  tenant_id, kb_id,
  slug: 'company-overview',
  patch: {section: 'leadership', content: 'CEO: Alice Wang (since 2020)',
          source_claim_id: claim.id, confidence: 0.98},
});

RAG side: merge patch into wiki_pages.body, re-lint, if pass → live; folded into next compile.

This feature launches 2026 Q2.

9.6 Shared Dashboard Metrics

Metric Source Meaning
AI Citation Rate (GEO) 7-dim scan % AI platforms mentioning brand
Fact Accuracy (GEO) NLI verify % mentions that are correct
Wiki Coverage (RAG) RAG compile stats % brand_facts covered by Wiki
CS Hit Rate (RAG) RAG query log % CS questions answerable
Repair Latency inject → next-scan verify Days until AI changes answer

Five axes quantifying “brand AI health.”

Key Takeaways

References

Navigation: ← Ch 8 · 📖 Contents · Ch 10 →

This site is open source. Improve this page.