> ## Documentation Index
> Fetch the complete documentation index at: https://docs.agentium.in/llms.txt
> Use this file to discover all available pages before exploring further.

# Build a support assistant

> One learning path from a model request to an application you can host.

Learn the parts of a support assistant through independent, runnable checkpoints: look up an order, continue a conversation, retrieve a policy, ask before a simulated refund, and serve a response. Each chapter isolates one integration so you can adapt it to your own application.

**You need:** TypeScript familiarity, a [matching SDK installation](/learn/installation), and an OpenAI API key for model and embedding calls. The browser execution explorer works without credentials.

## The path

| Chapter | What you add | How you know it works |
| - | - | - |
| [1. Answer a question](/learn/first-agent) | Instructions and a model | The assistant asks for an order ID |
| [2. Look up an order](/learn/tools) | A validated tool | ORD-1042 is reported as arriving Friday |
| [3. Remember the conversation](/learn/memory) | A stable session ID | A follow-up can name the earlier order |
| [4. Answer from policies](/learn/knowledge) | Retrieval with source IDs | The answer cites the return policy |
| [5. Ask before a refund](/learn/approval) | A human approval callback | Denial leaves the simulated refund count at zero |
| [6. Stream and inspect a run](/learn/streaming) | Text streaming and tool events | Text arrives incrementally; tool use is visible |
| [7. Serve an HTTP API](/learn/http) | An explicit local router | A curl request returns an agent response |
| [8. Prepare to ship](/learn/ship) | An application-specific deployment plan | Ownership, failure behavior, and evaluation have checks |

Each chapter is a standalone program, rather than the next revision of one cumulative application. Save its code as `support.ts` and replace the previous checkpoint. Chapters isolate the new capability so you can inspect it; the retrieval and approval checkpoints do not silently carry over every earlier tool. Once the pieces work, compose the tools your application needs.

The examples use a fictional order and an in-memory refund record. They never connect to a payment provider. Model wording varies, so verify the underlying tool result and state as well as the text.

## What is checked

The TypeScript examples are embedded from source files in this docs repository and checked against the sibling SDK's built public declarations. This checks API usage, not a provider's behavior or the quality of its answer. Run the verification prompts in your own environment before shipping.

<Card title="Begin with one request" icon="arrow-right" href="/learn/first-agent">Create the support Agent, inspect its result, and close it.</Card>


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.