Chatbot + Brando runtime flow

This example walks through an end-to-end flow:

1. A user asks a question in a support chatbot.
2. The chatbot’s backend:
   • resolves the relevant brando:Brand and brando:Context,
   • gathers Brando:VerbalIdentity and brando:Policy nodes,
   • optionally evaluates brando:AutomationRule nodes,
   • builds a runtime configuration for the LLM.
3. The LLM:
   • receives a system prompt derived from Brando,
   • generates an answer,
   • the answer is checked against Brando policies.

The goal is to show how the Brand Knowledge Graph becomes the operational control plane for a real chatbot.

---

1. Setup: Brando graph for a support chatbot

We’ll reuse a simplified version of the “Northstar Bank” example.

{
  "@context": {
    "schema": "https://schema.org/",
    "brando": "https://brandoschema.com/"
  },
  "@graph": [
    {
      "@id": "https://example.com/brand/northstar",
      "@type": "brando:Brand",
      "schema:name": "Northstar Bank",
      "brando:missionStatement": "Financial clarity with absolute trust.",
      "brando:coreValues": [
        "Clarity over complexity",
        "Do the right thing",
        "Long-term alignment with customers"
      ],
      "brando:hasContext": {
        "@id": "https://example.com/brand/northstar/context/retail-chatbot"
      },
      "brando:hasPolicy": {
        "@id": "https://example.com/brand/northstar/policies/global-brand-safety"
      }
    },
    {
      "@id": "https://example.com/brand/northstar/context/retail-chatbot",
      "@type": "brando:Context",
      "brando:audienceSegment": [
        "Retail banking customers",
        "Prospects researching savings accounts"
      ],
      "brando:audiencePersona": [
        "Everyday saver"
      ],
      "brando:domainContext": {
        "channel": "chatbot",
        "surface": "public-website",
        "region": "UK"
      },
      "brando:usesVerbal": {
        "@id": "https://example.com/brand/northstar/tokens/verbal/support"
      }
    },
    {
      "@id": "https://example.com/brand/northstar/tokens/verbal/support",
      "@type": "Brando:VerbalIdentity",
      "brando:toneOfVoice": "Reassuring and efficient. Focus on clear next steps.",
      "brando:dialogueStyle": "Use second person; acknowledge emotion, then move to options.",
      "brando:writingStyle": "Short paragraphs, concrete steps, minimal jargon.",
      "brando:approvedTerms": [
        "interest rate",
        "fees",
        "savings goals"
      ],
      "brando:prohibitedTerms": [
        "get rich quick",
        "guaranteed returns"
      ],
      "brando:mustDos": [
        "Offer at least one clear next step in each answer."
      ],
      "brando:mustNotDos": [
        "Minimise or dismiss user concerns."
      ]
    },
    {
      "@id": "https://example.com/brand/northstar/policies/global-brand-safety",
      "@type": "brando:Policy",
      "brando:enforcementLevel": "mandatory",
      "brando:riskTag": [
        "brand-safety",
        "financial-regulation"
      ],
      "brando:guardRails": [
        "Do not provide personalised investment advice.",
        "Do not promise guaranteed returns.",
        "Do not provide legal or tax advice."
      ],
      "brando:refusalStrategies": [
        "Explain that you cannot provide personalised advice and why.",
        "Offer general educational information instead.",
        "Suggest speaking to a regulated advisor for personal recommendations."
      ],
      "brando:retrievableInLLM": true
    }
  ]
}

This graph defines everything the chatbot needs to behave on-brand and within risk constraints.

---

2. Step-by-step runtime flow

2.1 User sends a message

Example:

“Hi, I’ve got £5,000 in savings – how should I invest it to get the best returns?”

The chatbot backend receives:

{
  "userId": "user-1234",
  "brandId": "https://example.com/brand/northstar",
  "contextId": "https://example.com/brand/northstar/context/retail-chatbot",
  "message": "Hi, I’ve got £5,000 in savings – how should I invest it to get the best returns?"
}

2.2 Brand OS resolves Brando nodes

The Brand OS (or brand middleware) uses brandId and contextId to:

1. Load the brando:Brand node.
2. Load the brando:Context node.
3. Follow brando:usesVerbal to the relevant Brando:VerbalIdentity.
4. Follow brando:hasPolicy to brando:Policy nodes.

Conceptual TypeScript:

const brand = getNode("https://example.com/brand/northstar");
const context = getNode("https://example.com/brand/northstar/context/retail-chatbot");

const Verbal = resolveLinkedNode(context, "brando:usesVerbal");
const policies = resolveLinkedNodes(brand, "brando:hasPolicy");

2.3 Assemble a runtime configuration

The Brand OS converts nodes into a RuntimeBrandConfig (domain model):

const runtimeConfig = {
  brand: {
    name: "Northstar Bank",
    mission: "Financial clarity with absolute trust.",
    coreValues: [
      "Clarity over complexity",
      "Do the right thing",
      "Long-term alignment with customers"
    ]
  },
  context: {
    channel: "chatbot",
    surface: "public-website",
    region: "UK",
    audienceSegment: [
      "Retail banking customers",
      "Prospects researching savings accounts"
    ],
    persona: "Everyday saver"
  },
  Verbal: {
    toneOfVoice: "Reassuring and efficient. Focus on clear next steps.",
    dialogueStyle: "Use second person; acknowledge emotion, then move to options.",
    writingStyle: "Short paragraphs, concrete steps, minimal jargon.",
    approvedTerms: ["interest rate", "fees", "savings goals"],
    prohibitedTerms: ["get rich quick", "guaranteed returns"],
    mustDos: ["Offer at least one clear next step in each answer."],
    mustNotDos: ["Minimise or dismiss user concerns."]
  },
  policies: [
    {
      enforcementLevel: "mandatory",
      riskTag: ["brand-safety", "financial-regulation"],
      guardRails: [
        "Do not provide personalised investment advice.",
        "Do not promise guaranteed returns.",
        "Do not provide legal or tax advice."
      ],
      refusalStrategies: [
        "Explain that you cannot provide personalised advice and why.",
        "Offer general educational information instead.",
        "Suggest speaking to a regulated advisor for personal recommendations."
      ]
    }
  ]
}

(How you map JSON-LD → these objects is implementation-specific, but follows the patterns in TypeScript model.)

---

3. Building the system prompt

The LLM adapter converts runtimeConfig into a system prompt. For example:

You are Northstar Bank's public website chatbot.

Brand identity:
- Mission: Financial clarity with absolute trust.
- Core values:
  - Clarity over complexity
  - Do the right thing
  - Long-term alignment with customers

Context:
- Channel: Chatbot on the public website (UK region).
- Audience: Retail banking customers and prospects researching savings accounts.
- Persona: Everyday saver (non-expert; may feel anxious about money).

Tone of voice:
- Reassuring and efficient. Focus on clear next steps.
- Use second person; acknowledge emotion, then move to options.
- Short paragraphs, concrete steps, minimal jargon.

Language rules:
- Approved terms (prefer these): interest rate; fees; savings goals
- Prohibited terms (never use): get rich quick; guaranteed returns

Behavioural rules:
- Always offer at least one clear next step in each answer.
- Never minimise or dismiss user concerns.

Brand and risk guard rails (MANDATORY):
- Do not provide personalised investment advice.
- Do not promise guaranteed returns.
- Do not provide legal or tax advice.

Refusal patterns:
- If a user asks for personalised advice, explain that you cannot provide it and why.
- Offer general educational information instead.
- Suggest speaking to a regulated advisor for personal recommendations.

Your responses must follow all of these rules.

The model then receives:

• this system message,
• plus the user’s message and any other context.

---

4. LLM response and policy checking

4.1 Example LLM output (ideal)

“I can’t provide personalised investment advice, but I can walk you through some general options. For example, you might consider how much you want to keep as easy-access savings versus longer-term goals. We can look at typical interest rates, any fees, and how these might support your savings goals. A next step is to review our savings products page and, if you’d like personalised guidance, speak to a regulated advisor.”

This response is:

• on tone,
• uses approved terms,
• refuses personalised advice appropriately,
• suggests a clear next step.

4.2 Policy enforcement (non-normative)

Your Brand OS can apply additional checks post-generation using the same Brando properties:

• Ensure no prohibited terms appear (brando:prohibitedTerms).
• Ensure mandatory elements are present (brando:mustDos).
• Flag or block responses that violate brando:guardRails.

For example:

import { evaluateResponse } from "./policy-engine";

const result = evaluateResponse({
  text: modelOutput,
  policies: runtimeConfig.policies,
  Verbal: runtimeConfig.Verbal
});

if (!result.compliant) {
  // Option A: ask the model to self-correct with explicit feedback
  // Option B: provide a fallback “safe” response
  // Option C: route to human review for high-risk cases
}

The policy engine is implementation-specific but uses Brando properties as inputs.

---

5. Where automation rules fit

If you also define brando:AutomationRule nodes, they can adjust behaviour based on metrics.

Example rule (simplified):

{
  "@id": "https://example.com/brand/northstar/automation/unsafe-output-monitor",
  "@type": "brando:AutomationRule",
  "brando:appliesTo": "{\"contexts\":[\"retail-chatbot\"],\"region\":\"UK\"}",
  "brando:triggerType": "metric-threshold",
  "brando:monitoredMetric": "{\"name\":\"unsafe_output_rate\",\"threshold\":0.02,\"operator\":\">\",\"window\":\"30d\"}",
  "brando:dataSource": "https://monitoring.example.com/api/metrics/unsafe_output_rate",
  "brando:automationAction": "{\"actionType\":\"increase_enforcement\",\"targetPolicies\":[\"https://example.com/brand/northstar/policies/global-brand-safety\"],\"newLevel\":\"mandatory\",\"notify\":[\"risk-team@northstar.example\"]}",
  "brando:enforcementLevel": "mandatory",
  "brando:riskTag": [
    "brand-safety",
    "llm-guardrails"
  ]
}

Your Brand OS might:

1. Periodically evaluate this rule.

2. If unsafe outputs exceed 2% in 30 days:

3. make certain policies mandatory for all chatbot interactions,

4. enable extra checks,
5. notify the risk team.

This is how an AI-managed Brand Knowledge Graph can tighten controls as risk increases.

---

6. Summary

In this chatbot example, Brando vocab v1.3:

• provides a structured definition of:

• Brand (identity, mission, values),

• Context (channel, audience, region),
• Verbal expression (tone, style, vocabulary),
• Policies (guard rails, refusal strategies, enforcement).
• allows your Brand OS to assemble runtime configurations for each interaction.
• lets LLMs and policy engines work from the same source of truth, instead of ad-hoc prompts.

For more detail:

• Brand Operating System architecture
• Runtime Integration (LLMs, MCP, APIs)
• Minimal brand graph
• House of brands example