F10: Multi-Language AI Output

F10: Multi-Language AI Output

Why This Exists

Standard GRC serves Brazilian and international enterprises. A CISO at a São Paulo SaaS company needs board reports in Portuguese, while their ISO 27001 auditor from Deloitte needs the same findings in English. Without native multi-language support, customers would need to manually translate 300+ control findings — defeating the purpose of AI-driven automation.

The solution: a locale parameter threaded from the API layer through the agent executor into the LLM system prompt, ensuring all AI output respects the requested language.

---

Architecture

sequenceDiagram
    autonumber
    participant Client as API Client
    participant Gateway as API Gateway
    participant Executor as Agent Executor
    participant LLM as GPT-4o

    Client->>Gateway: POST /agent-runs?locale=en
    Gateway->>Executor: execute(run, context { locale: "en" })
    Executor->>Executor: Build system prompt with locale
    Executor->>LLM: System: "Output Language: English (en)"
    LLM-->>Executor: JSON findings in English
    Executor-->>Gateway: AgentRunResponse
    Gateway-->>Client: 200 { findings in English }

flowchart LR
    subgraph "Locale Resolution"
        A["?locale=en"] --> B["context.locale = 'en'"]
        C["No param"] --> D["context.locale = 'pt-BR'"]
    end
    B --> E["System Prompt Injection"]
    D --> E
    E --> F["LLM generates in target language"]

    style A fill:#2d333b,stroke:#6d5dfc,color:#e6edf3
    style B fill:#2d333b,stroke:#6d5dfc,color:#e6edf3
    style C fill:#2d333b,stroke:#6d5dfc,color:#e6edf3
    style D fill:#2d333b,stroke:#6d5dfc,color:#e6edf3
    style E fill:#2d333b,stroke:#6d5dfc,color:#e6edf3
    style F fill:#2d333b,stroke:#6d5dfc,color:#e6edf3

---

Implementation

Locale in Agent Runtime Context

The locale field is added to AgentRuntimeContextSchema in packages/schemas:

locale: z.enum(["pt-BR", "en"]).default("pt-BR")

(packages/schemas/src/agent-runtime.ts)

System Prompt Injection

The executor resolves the locale and injects it into the LLM system prompt:

const locale = context.locale ?? "pt-BR";
const localeName = locale === "pt-BR" ? "Brazilian Portuguese" : "English";

const systemPrompt = `You are the ${contract.display_name}.
Responsibility: ${contract.responsibility}
Output Language: ${localeName} (${locale}). Write ALL findings, summaries,
recommendations, and field values in ${localeName}.`;

(packages/agent-runtime/src/executor.ts:97-103)

API Route Threading

The API route reads ?locale= from the query string and injects it into the agent execution context:

// In agent-runtime.routes.ts
const locale = url.searchParams.get("locale") ?? "pt-BR";
context.locale = locale;

(apps/api-gateway/src/routes/agent-runtime.routes.ts)

Report Renderer Labels

The Markdown renderer uses a bilingual label map:

const LABELS = {
  "en": { summary: "Executive Summary", findings: "Findings", ... },
  "pt-BR": { summary: "Resumo Executivo", findings: "Achados", ... }
};

(packages/reporting/src/renderers/markdown-renderer.ts)

---

Supported Locales

Locale	Language	Default?
pt-BR	Brazilian Portuguese	✅ Yes
en	English	No

---

API Usage

# English output
curl -X POST "$BASE/assessments/$ID/agent-runs?locale=en" \
  -H "Authorization: Bearer $API_KEY"

# Portuguese (default)
curl -X POST "$BASE/assessments/$ID/agent-runs" \
  -H "Authorization: Bearer $API_KEY"

SDK

const run = await client.agents.execute(assessmentId, {
  agent_id: "evidence-evaluator",
  locale: "en",
  input: { controlRequirement: "...", evidenceDescription: "..." }
});
// run.output.auditor_notes → in English

---

Affected Components

Component	File	Change
Schemas	packages/schemas/src/agent-runtime.ts	Added locale field
Executor	packages/agent-runtime/src/executor.ts	System prompt injection
API Routes	apps/api-gateway/src/routes/agent-runtime.routes.ts	Query param threading
Renderer	packages/reporting/src/renderers/markdown-renderer.ts	Bilingual labels

---

References

• executor.ts — Locale resolution and prompt injection
• agent-runtime.routes.ts — API query param handling
• Product Discovery — F10 RICE score: 600 (highest)