# Free, private AI summaries in 15 minutes

In this tutorial you'll give SearchBox a local AI assistant: install Ollama, download a small AI model, connect SearchBox to it, and run your first search that comes back with a written answer — citations included. Everything you set up here runs on your own computer and costs nothing.

If you want the background on what the AI Summary panel does before diving in, read [AI summaries](../ai-summaries.md) first.

> 🔒 Nothing in this tutorial creates an account, asks for a credit card, or sends your documents anywhere. The one and only internet download is the AI model itself — after that, the whole feature works offline.

## Step 1: Install Ollama (about 5 minutes)

Ollama is a free program that runs AI models on your own machine. SearchBox talks to it locally to write summaries.

1. Go to [ollama.com](https://ollama.com) in your browser.
2. Click **Download** and pick your operating system (Windows, macOS, or Linux).
3. Run the installer and follow its prompts.
4. When it finishes, Ollama starts running quietly in the background — on Windows you'll see a small llama icon in the system tray, near the clock.

![The Ollama download page](../img/ollama-download-page.png)
<!-- SCREENSHOT NEEDED: Open https://ollama.com in a browser. Capture the landing page showing the Download button and the Ollama llama logo. Crop to the main hero area of the page. -->

## Step 2: Download a model (about 5 minutes, depending on your internet)

An AI model is the "brain" Ollama runs — a single large file you download once. SearchBox doesn't require a specific one; you type whichever model's name into its settings later. A good starting point is **llama3.2:3b** — small (around 2 GB), quick, and capable enough for summarizing documents.

1. Open a terminal. On Windows: press the Start button, type `powershell`, and press **Enter**.
2. Type the following and press **Enter**:

   ```
   ollama pull llama3.2:3b
   ```

3. Wait for the download bar to reach 100%. This is the one big download in the whole setup.
4. Confirm it worked by typing:

   ```
   ollama list
   ```

   You should see `llama3.2:3b` in the list.

![PowerShell after pulling the model, with ollama list showing it installed](../img/ollama-pull-terminal.png)
<!-- SCREENSHOT NEEDED: In PowerShell, run `ollama pull llama3.2:3b` and let it complete, then run `ollama list`. Capture the window showing the completed pull (100% / success) and the list output containing llama3.2:3b with its size. -->

> 💡 **Tip:** Have a powerful computer (16 GB+ of memory)? You can pull a larger model later for better answers — anything in Ollama's library works. Whatever you pull, its exact name goes into SearchBox's **Model** field in the next step.

## Step 3: Connect SearchBox to Ollama (2 minutes)

1. Open SearchBox and click the gear icon (**Settings**).
2. Scroll to the **AI Search** section.
3. Turn on the **Enable AI Search** toggle. An **Ollama Configuration** area appears below it.
4. Leave **Server URL** as `http://localhost:11434` — that's Ollama's standard local address. ("localhost" means "this computer" — the connection never touches the internet.)
5. In the **Model** field, type exactly:

   ```
   llama3.2:3b
   ```

6. Click **Test Connection** in the **Connection Status** row. The status dot turns green and the text reads **Connected**.
7. Check the **Available Models** row — `llama3.2:3b` should be listed. If not, click **Refresh**.

![SearchBox Settings with AI Search enabled and Ollama connected](../img/settings-ai-search.png)
<!-- SCREENSHOT NEEDED: In SearchBox Settings, AI Search section: Enable AI Search toggle ON, Model field containing llama3.2:3b, Connection Status showing a green dot with "Connected", and Available Models listing llama3.2:3b. Capture the whole AI Search section. -->

> ⚠️ If the status says **Not connected — is Ollama running?**, make sure Ollama is actually running (on Windows, look for the llama icon in the system tray, or start Ollama from the Start menu), then click **Test Connection** again.

## Step 4: Run a search that gets a summary

1. Go back to SearchBox's home page.
2. Type a question or topic you know your indexed documents cover — a topic that spans a few files works best, for example `insurance renewal` or `project timeline`.
3. Press **Enter** or click **Search Documents**.
4. Your results appear first, as always. A moment later the **AI Summary** panel appears above them and the answer streams in word by word.

When it finishes, look around the panel: the **high / medium / low** confidence badge next to the title, the **Top Sources** cards showing the documents the answer leaned on most, and the footer with the model name and generation time. The very first summary after a reboot is the slowest — the model is loading into memory. After that, expect a handful of seconds.

![Your first AI summary above the search results](../img/ai-summary.png)
<!-- SCREENSHOT NEEDED: On the results page, capture a completed AI Summary panel for a real query: 🤖 AI Summary header with confidence badge, streamed answer text containing [1] [2] citation links, Top Sources cards at the side, and footer metadata (sources count, seconds, model name). Include a search result or two below the panel. -->

> 💡 **Tip:** Search the same thing again and the answer appears instantly with a **⚡ Cached** tag — SearchBox saved it. The circular-arrow **Regenerate summary** button in the panel's corner always writes a fresh one.

## Step 5: Click a citation

The numbers like **[1]** and **[2]** inside the answer aren't decoration — each is a link to the document that claim came from.

1. Hover over a citation to preview the file's name, type, and size.
2. Click it. SearchBox opens that document in the viewer, with your search carried along, so you can verify the AI's claim against the actual text.

![The document viewer opened from a citation link](../img/citation-open-document.png)
<!-- SCREENSHOT NEEDED: From a finished AI summary, click a [1] citation link. Capture the document viewer page that opens, showing the document content. Ideally show the search terms highlighted in the document text. -->

That habit — read the summary, click the citation, confirm the source — is the whole workflow. The AI gives you the head start; your documents remain the truth.

## Is it actually private?

Yes, and you can verify it yourself rather than take anyone's word:

- **The AI lives at `localhost`.** SearchBox sends your search phrase and short snippets from your top results to Ollama at `http://localhost:11434` — an address that, by definition, is your own computer. SearchBox itself also only listens on your own computer.
- **No account, no key.** You never logged into anything during this tutorial. There is no service on the other end to keep logs.
- **The offline test.** Once the model is downloaded, turn off Wi-Fi and run a search. The summary still generates — because nothing about it ever needed the internet.

The only time Ollama touches the network is when *you* ask it to download a model.

## Done

You now have private, free AI summaries on every search. If you ever want them gone, flip off **Enable AI Search** in **Settings** — details in [AI summaries](../ai-summaries.md).

- [AI summaries](../ai-summaries.md) — full reference for the summary panel and its settings
- [Searching your documents](../searching.md) — filters, syntax, and getting more out of results

---
[← Back to the manual](../README.md)