Chat plugin

Single- or multi-turn chat completions against any Ollama-compatible endpoint.

Stable Reading time ~ 2 min Edit on GitHub

The Chat plugin is the simplest piece of the AI module. It wraps Ollama’s /api/chat with a fluent builder, conversation memory, automatic fallback to a backup model, and a one-shot ask() shortcut.

API at a glance

$ai   = new \Nibiru\Module\Ai\Ai();
$chat = $ai->chat();

$chat->system('Be terse.');           / optional system prompt
$chat->model('qwen2.5-coder:14b');    / override the configured model
$chat->temperature(0.2);              / override config
$chat->maxTokens(512);                / override config

$chat->user('Hello');                 / append a user message
$chat->assistant('Hi.');              / append an assistant message (rare)

$reply = $chat->complete();           / run the call, return text
$reply = $chat->ask('How are you?');  / = ->user(...)->complete()

$chat->reset();                       / clear messages, keep model + system
$chat->history();                     / [{role, content}, …]

One-shot

echo (new \Nibiru\Module\Ai\Ai())
    ->chat()
    ->ask('In one sentence, what does Form::create() do?');

Multi-turn

$chat = $ai->chat();

$chat->user('Name three Nibiru singletons.');
$singletons = $chat->complete();        / appended to history

$chat->user('What does the second one do?');
$detail = $chat->complete();            / model has full context

Overriding model + style per call

$german = $ai->chat()
    ->system('Answer in German. Be precise.')
    ->model('qwen2.5-coder:14b')
    ->temperature(0.1)
    ->ask('Wie definiere ich einen Controller?');

Automatic fallback

If chat.model (e.g. nibiru-coder:1.0) isn’t on the Ollama server, the plugin re-runs the call with chat.fallback_model (e.g. qwen2.5-coder:14b). This keeps your app working while you’re still building the fine-tune.

[AI]
chat.model          = "nibiru-coder:1.0"
chat.fallback_model = "qwen2.5-coder:14b"

Switching providers

Default: Ollama. To use Anthropic Claude as the backend:

[AI]
chat.provider       = "anthropic"
anthropic.api_key   = "sk-ant-..."
anthropic.model     = "claude-haiku-4-5-20251001"

The Chat plugin doesn’t ship Anthropic transport in the framework module yet — for now, the docs site’s scripts/lib/providers.mjs pattern is the reference. (See the Roadmap.)

A practical pattern: chat-as-an-action

namespace Nibiru;
use Nibiru\Adapter\Controller;
use Nibiru\Module\Ai\Ai;

class supportController extends Controller
{
    public function askAction(): void {
        View::forwardToJsonHeader();
        $q = trim($this->getPost('question', ''));
        if ($q === '') {
            View::assign(['data' => ['error' => 'question required']]);
            return;
        }
        $reply = (new Ai())->chat()
            ->system('You are the Nibiru support assistant. Be brief.')
            ->ask($q);
        View::assign(['data' => ['answer' => $reply]]);
    }

    public function pageAction(): void {}
    public function navigationAction(): void {
        JsonNavigation::getInstance()->loadJsonNavigationArray();
    }
}

Six lines plus boilerplate, and you’ve got an AJAX-callable AI endpoint backed by your own Ollama.

Common pitfalls

Forgetting complete(). The fluent builder doesn’t run anything until complete() or ask() is called.
Calling assistant() between user turns. It’s there for replaying a saved conversation, not for normal use.
Long conversations. Each turn re-sends the full history. Trim with reset() or by slicing history() when you no longer need older context.
Setting temperature too low. 0 makes the model rigid; 0.3–0.5 is the sweet spot for technical answers.