Example Anatomy

An example should show enough of the application flow to make Anvia’s role clear. It should not stop at a single isolated method call, and it should not become a full app tutorial.

Required Shape

Part	What it should show
Build goal	The product or workflow someone is trying to ship.
Flow	The steps from request or source data to the final product result.
Anvia surface	The agents, tools, models, embeddings, vector stores, extractors, pipelines, observers, or adapters involved.
Application boundary	The code your app owns: auth, services, storage, permissions, queues, jobs, and audit.
Minimal TypeScript	One compact vertical slice that connects the important pieces.
Swap points	Which provider, vector store, model, or transport can change without rewriting the flow.
Production checks	The failure modes that matter before real users or real data.

Good Scope

For a RAG example, a single embedTexts(...) call is too small. The useful scope is source loading, OCR or text extraction, chunking, embedding, vector storage, retrieval, and grounded answer behavior.

For a tool example, a createTool(...) call with a schema is too small. The useful scope is request-scoped user and tenant data, service-level authorization, output narrowing, audit, and denied paths.

For an agent example, new AgentBuilder(...).build() is too small. The useful scope is the route or job boundary, runner, scoped tools, context, model call, trace metadata, persistence, and response shape.

Compact Vertical Slice

The code can still be short:

export async function runWorkflow(input: WorkflowInput) {
  const user = await input.auth.requireUser();
  const agent = createWorkflowAgent({
    model: input.model,
    user,
    services: input.services,
    knowledge: input.knowledge,
  });

  const response = await agent
    .prompt(input.message)
    .withTrace({
      name: "workflow",
      userId: user.id,
      metadata: { tenantId: user.tenantId },
    })
    .send();

  await input.events.record({
    userId: user.id,
    output: response.output,
    traceId: response.trace?.traceId,
  });

  return { output: response.output };
}

The names can be product-specific. The pattern should stay visible: resolve app state, compose Anvia runtime, run the workflow, persist what matters, and return a product result.

What To Avoid

• examples that only call one API and never show where it belongs
• examples that hide auth, permissions, or persistence in comments
• examples that put user or tenant ids in model-controlled arguments
• examples that show RAG without ingestion or retrieval boundaries
• examples that show side effects without idempotency, approval, or audit
• examples that copy package reference docs instead of connecting features into a build pattern