Build Your First AI System

Most organizations start with prompts.

Some move to agents.

A few build workflows.

But production AI requires more than any individual component.

It requires AI Systems.

In this guide, we’ll build an Employee IT Support Assistant that can:

• Understand employee requests
• Retrieve company policies
• Interact with internal systems
• Request approvals
• Execute business actions
• Provide complete auditability

By the end, you’ll have built a production-ready AI system.

---

Video Tutorial

Follow the complete walkthrough to create your first AI System in Karyam.

---

The Problem

An employee submits a request:

My VPN access has expired. Can you restore it?

A typical chatbot might answer:

Please contact IT support.

A production AI system should do much more.

---

The Desired Outcome

Our AI system will:

1. Understand the employee request.
2. Retrieve the company’s VPN policy.
3. Check employee status in the HR system.
4. Determine whether approval is required.
5. Request manager approval.
6. Create an IT support ticket.
7. Record the complete execution history.

---

The Final Architecture

Employee Request
        ↓
AI Agent
        ↓
Vector Search
        ↓
Skills & MCP Tools
        ↓
Approval Workflow
        ↓
IT Ticket Creation
        ↓
Observability & Audit Logs

This is an AI System.

Not a chatbot.

---

Step 1 — Configure Enterprise Retrieval

Every AI system starts with context.

Before an agent can understand company policies, we need to configure the retrieval pipeline.

In Karyam, enterprise knowledge is powered by:

Vector Database
        +
Embedding Model
        +
Embedding Listener
        +
File Upload
        ↓
RAG Pipeline

---

Configure an Embedding Model

Navigate to:

Models → Create Model

Select a model of type:

Embedding

Examples:

• text-embedding-3-large
• Gemini Embedding
• Voyage AI
• Ollama Embeddings

This model converts uploaded content into vectors.

---

Create a Vector Database

Navigate to:

Vector Databases → Create Vector Database

Choose your preferred provider from the list

This database stores embeddings generated from uploaded files.

---

Create an Embedding Listener

Navigate to:

Listeners → Create Listener

Configure:

Setting	Value
Type	Embedding Listener
Folder Path	/knowledge-base/it-support
Vector Database	IT Support DB
Embedding Model	text-embedding-3-large

Embedding listeners automatically react to new file uploads.

---

Upload Enterprise Documents

Upload:

• VPN Policy.pdf
• IT Access SOP.pdf
• Employee Access Guidelines.pdf

Example:

knowledge-base/
└── it-support/
    ├── vpn-policy.pdf
    ├── access-sop.pdf
    └── employee-guidelines.pdf

The ingestion pipeline executes automatically:

File Uploaded
      ↓
Listener Triggered
      ↓
Parsing & Chunking
      ↓
Embedding Generation
      ↓
Vector Storage
      ↓
RAG Run Created

No manual synchronization is required.

---

Verify Ingestion

Navigate to:

RAG Runs

You can observe:

• File parsing
• Chunk generation
• Embedding creation
• Vector storage
• Errors and retries

Every ingestion operation is fully observable.

---

Step 2 — Create an AI Agent

Navigate to:

Agents → Create Agent

Configure:

Setting	Value
Name	IT Support Assistant
Model	Claude Sonnet
Temperature	0.2
Max Steps	10

Attach:

• HR Skills
• Ticketing Skills
• MCP Tools
• Approval Flow

---

System Instructions

You are an IT Support Assistant responsible for handling VPN access requests.

Responsibilities:
- Verify user eligibility.
- Retrieve company policies.
- Request approvals when required.
- Create IT tickets after approval.

Rules:
- Never bypass approval requirements.
- Always reference retrieved company knowledge.
- Escalate uncertain situations to human operators.

---

Step 3 — Connect Skills and MCP Tools

AI becomes useful when it can act.

Attach:

Skills

• HR System Lookup
• Ticket Creation
• Email Notifications

MCP Tools

• HR Database
• Active Directory
• Jira Service Management

Your AI system can now interact securely with business systems.

---

Step 4 — Build the Workflow

Navigate to:

AI Flows → Create Flow

Create:

Employee Request
        ↓
Retrieve Context
        ↓
Check HR Status
        ↓
Approval Required?
      ↙       ↘
    Yes        No
     ↓          ↓
Manager      Create Ticket
Approval         ↓
     ↓            ↓
Create Ticket  Notify Employee

Business processes require control.

Karyam workflows combine intelligence with governance.

---

Step 5 — Add Human Approval

Configure an approval step:

Trigger	Approval Required
VPN Access Restoration	Yes
Administrator Access	Yes
Password Reset	No

Approvals can be routed to:

• Managers
• IT Teams
• Security Teams
• Compliance Teams

This is Human-in-the-Loop (HITL).

---

Step 6 — Execute the System

The employee asks:

My VPN access has expired. Can you restore it?

The AI system executes:

Understand Request
        ↓
Retrieve VPN Policy
        ↓
Check HR Status
        ↓
Request Approval
        ↓
Create Ticket
        ↓
Notify Employee

---

Step 7 — Observe Everything

Production AI requires visibility.

Navigate to:

Runs
RAG Runs
Audit Logs

Monitor:

• Agent Logs
• Workflow Logs
• RAG Runs
• Tool Executions
• Approval Events
• Audit Logs
• Token Usage
• Costs

---

Example Execution Trace

10:01 Employee Request Received
10:01 VPN Policy Retrieved
10:02 Employee Status Verified
10:02 Approval Request Sent
10:05 Manager Approved
10:05 Ticket Created
10:05 Employee Notified

Every action is recorded.

Every decision is traceable.

Every workflow is observable.

---

The Karyam Model

Vector Databases
        +
Embedding Models
        +
Embedding Listeners
        +
Agents
        +
Skills
        +
Flows
        +
Approvals
        +
Observability
            ↓
Production AI Systems

This is the Karyam approach.

Build AI Systems. Not Hacks.

---

Next Steps

Congratulations 🎉

You have built your first AI System in Karyam.

Next you can explore:

• Multi-agent collaboration
• Advanced workflows
• Custom MCP integrations
• Enterprise governance
• On-prem deployments