@piatrajs/ai

Thin wrapper over the official openai SDK, preset for the Piatra AI Gateway. It does exactly three things: presets the gateway base URL, reads the key from PIATRA_AI_KEY, and exports typed piatra.* model-id constants. Everything else — chat, streaming, retries — is the OpenAI SDK's, unchanged.

Install

npm install @piatrajs/ai

Usage

import PiatraAI, { MODELS } from '@piatrajs/ai';

const client = new PiatraAI();
// apiKey  defaults to process.env.PIATRA_AI_KEY
// baseURL defaults to https://ai.piatra.center/v1

const res = await client.chat.completions.create({
  model: MODELS.GLM, // 'piatra.glm'
  messages: [{ role: 'user', content: 'What is default logic?' }],
});

Streaming and multimodal input are the OpenAI SDK's, unchanged:

const stream = await client.chat.completions.create({
  model: MODELS.GLM,
  messages,
  stream: true,
});
for await (const chunk of stream) {
  /* … */
}

await client.chat.completions.create({
  model: MODELS.MEDGEMMA, // the vision-capable model
  messages: [{
    role: 'user',
    content: [
      { type: 'text', text: 'describe this radiograph' },
      { type: 'image_url', image_url: { url: dataUrl } },
    ],
  }],
});

Configuration

| Option | Default | Notes | | --------- | ---------------------------------- | --------------------------------------------------------------------- | | apiKey | process.env.PIATRA_AI_KEY | The pi_live_* key, issued and owned by core. | | baseURL | https://ai.piatra.center/v1 | Or set PIATRA_AI_BASE_URL. In-cluster callers use the internal DNS. |

Any other OpenAI client option is accepted and passed through unchanged.