Function Calling — Complete Guide

Introduction

Function Calling — Complete Guide is essential for developers building Enterprise Agentic AI Platform — Toolliyo's 100-article ASP.NET Core with Agentic AI path covering AI foundations, ASP.NET Core AI APIs, Semantic Kernel, multi-agent orchestration, RAG/vector stores, AI security, cloud-native deployment, SaaS patterns, and enterprise projects (CRM copilot, hospital assistant, ERP, analytics, coding assistant).

Teams shipping production copilots need function calling with tool calling, tenant isolation, eval harnesses, and observability — not single-turn chat demos.

After this article you will

• Explain Function Calling for ASP.NET Core agentic AI — orchestration, tools, memory, and governance
• Apply function calling inside Enterprise Agentic AI Platform (Hospital Assistant)
• Compare chatbot demos vs production multi-agent systems with eval and observability
• Answer interview questions on Semantic Kernel, RAG, tool calling, and AI security
• Connect to Article 35 in the 100-lesson path

Prerequisites

• Software: .NET 8 SDK, VS 2022, Azure OpenAI access
• Knowledge: ASP.NET Core, Prompt Engineering, Agentic AI with .NET
• Previous: Article 33 — Tool Calling — Complete Guide
• Time: 24 min reading + 30–45 min hands-on

Concept deep-dive

Level 1 — Analogy

Tool calling is handing the model a remote control — it picks which API button to press; you execute server-side with RBAC.

Level 2 — Technical

Function Calling orchestrates specialist agents — triage/research/writer roles, message handoffs, idempotent tool calls, and supervisor gates on Hospital Assistant.

Level 3 — Agent orchestration flow

[User / Partner API / Webhook]
       ▼
[ASP.NET Core Agent Orchestrator API]
       ▼
[Semantic Kernel + Plugins / Tool Schemas]
       ▼
[Specialist Agents + Memory (Redis / Qdrant / pgvector)]
       ▼
[Policy · Approval · OpenTelemetry · Audit Log]
       ▼
[Project module: Hospital Assistant]

Solution structure

EnterpriseAgenticAI/
├── src/
│   ├── EnterpriseAgenticAI.Api/       ← ASP.NET Core agent endpoints
│   ├── EnterpriseAgenticAI.Agents/    ← SK agents, plugins, orchestrators
│   ├── EnterpriseAgenticAI.Core/      ← Tool schemas, policies, contracts
│   └── EnterpriseAgenticAI.Tests/     ← Golden-task eval + xUnit
├── infra/
│   ├── docker-compose.yml             ← API + Qdrant + Redis
│   └── k8s/                           ← AKS deployment
└── eval/                              ← Golden scenarios per agent version

Hands-on implementation — Hospital Assistant

Build Function Calling in Enterprise Agentic AI Platform for Hospital Assistant: ASP.NET Core API + Semantic Kernel, tool calling with RBAC, RAG memory, and golden-task eval before deploy.

1. Open EnterpriseAgenticAI solution (Api, Agents, Core, Tests).
2. Configure ASP.NET Core AI endpoints, DI, and Azure OpenAI in Program.cs.
3. Register Semantic Kernel plugins and tool schemas with read-only defaults.
4. Add RAG memory (Qdrant/pgvector) with tenant-isolated namespaces.
5. Run golden-task xUnit eval and export OpenTelemetry traces per agent step.

Anti-pattern (OpenAI in controller, no auth, no eval, shared tenant memory)

// ❌ BAD — OpenAI in controller, no auth, no eval
[HttpPost("chat")]
public async Task<string> Chat(string msg)
{
    var client = new OpenAIClient(apiKeyFromAppsettings); // secret in git
    return await client.GetChatCompletion(msg); // no RAG, no audit, no tenant isolation
}

Production-style ASP.NET Core agent API

// ✅ PRODUCTION — Function Calling (Hospital Assistant)
public class AgentOrchestrator : IAgentOrchestrator
{
    public async Task<AgentResult> RunAsync(AgentRequest req, CancellationToken ct)
    {
        var kernel = _kernelFactory.Create(req.TenantId);
        kernel.ImportPluginFromObject(new CrmReadPlugin(_db), "crm");

        using var activity = ActivitySource.StartActivity("CrmCopilot");
        var agent = _agentFactory.Create("CrmCopilot", kernel, readOnlyTools: true);
        var result = await agent.InvokeAsync(req.Message, ct);
        await _audit.LogAsync(req, result, activity);
        return result;
    }
}

Complete example

// Function Calling — Enterprise Agentic AI Platform (Hospital Assistant)
builder.Services.AddScoped<IAgentOrchestrator, AgentOrchestrator>();

Enterprise agent examples

HIPAA hospital assistant

Appointment booking, report summarization — PHI encrypted, break-glass access, immutable audit.

Hospital Assistant triage agent

Multi-agent: triage → clinical KB RAG → writer → supervisor approval before patient-facing text.

Enterprise Agentic AI Platform — Hospital Assistant · Article 34

Evaluating agent workflows

[Fact]
public async Task Agent_PassesGoldenTasks()
{
    var result = await _agentEval.RunGoldenTasksAsync("hospital-assistant-v1");
    Assert.True(result.SuccessRate >= 0.85);
}

Common errors & fixes

• Fat controllers calling OpenAI directly with secrets in appsettings — Agent orchestrator service + Key Vault; thin minimal APIs with auth and rate limits.
• Streaming responses without cancellation tokens — Use IAsyncEnumerable, HttpContext.RequestAborted, and timeout middleware on long agent runs.
• RAG without tenant ACL on vector chunks — Filter retrieval by tenant_id; encrypt embeddings at rest for regulated workloads.
• No token cost or latency dashboards — OpenTelemetry spans per SK step; Application Insights metrics for cost per tenant/day.

Best practices

• 🟢 Version agent configs; gate deploy on golden-task eval
• 🟢 Sandbox tools — read-only by default; approval for writes
• 🟡 Start with one specialist agent + RAG before multi-agent
• 🟡 Cap planner iterations; log OpenTelemetry spans per tool call
• 🔴 Never deploy without tenant-isolated memory and audit logs
• 🔴 Never trust user input as system instructions — use delimiters

Interview questions

Mid level

Q1: How does Function Calling fit an ASP.NET Core agent architecture?
A: Orchestrator API → SK Kernel → plugins/tools → memory → auth/audit → OpenTelemetry for Hospital Assistant.

Q2: Semantic Kernel vs calling OpenAI directly?
A: SK provides plugins, planners, DI integration, and testable abstractions — raw API needs manual wiring.

Q3: How do you secure tool calling?
A: JSON schemas, RBAC allowlists, read-only defaults, human approval for writes, audit every invocation.

Senior / architect level

Q4: Single agent vs multi-agent?
A: Single for focused tasks; multi-agent when triage/research/write/supervise improves quality and safety.

Q5: How do you evaluate agents before deploy?
A: Golden-task xUnit suite — success rate, latency, token cost; block deploy on regression.

Q6: Multi-tenant RAG isolation?
A: Separate vector namespaces or tenant_id filters; encrypt at rest; never mix customer documents in retrieval.

Summary & next steps

• Article 34: Function Calling — Complete Guide
• Module: Module 4: AI Agents and Multi-Agent Systems · Level: INTERMEDIATE
• Project module: Hospital Assistant

Previous: Tool Calling — Complete Guide
Next: Agent Communication — Complete Guide

Practice: Add one SK plugin with golden-task test — commit with feat(aspnet-agentic-ai): article-034.

FAQ