AI extraction

Zyte's AI outputs (product, productList, productNavigation, article, articleList, articleNavigation, forumThread, jobPosting, jobPostingNavigation, pageContent, serp and customAttributes) run through any LLM you choose:

import { llmExtractors, openAICompatible, Xyte } from '@xtandard/xyte';

const xyte = new Xyte({
  extractors: llmExtractors({
    complete: openAICompatible({ baseURL: 'https://openrouter.ai/api/v1', apiKey: process.env.OPENROUTER_API_KEY, model: 'openai/gpt-5-mini' }),
  }),
});

const { product } = await xyte.extract({ url: 'https://shop.example/p/123', product: true });
// { name, price, currency, sku, images, …, url, metadata: { dateDownloaded, probability } }

How it works:

1. The page is downloaded as for any request: HTTP by default, or the browser when productOptions.extractFrom is browserHtml or the request also asks for browserHtml.
2. It's converted to Markdown with the built-in converter: main content for articles, jobs and forum threads, the full page for products and listings. It's cut to maxInputChars (default 60 000, ~15k tokens).
3. The model gets the page plus Zyte's own output schema for the field, bundled from the OpenAPI spec (OUTPUT_SCHEMAS, regenerated with bun run gen:types). The response therefore has Zyte's field names and types.
4. metadata.probability is the model's estimate that the page really is of that type (0 when nothing was found), and metadata.dateDownloaded is set by xyte.

customAttributes uses the request's own schema and returns { values, metadata: { inputTokens, outputTokens, textInputTokens } }, as on Zyte. customAttributesOptions.maxInputTokens caps the page text.

Extracting from your own HTML

extractFrom: "userHtml" in a field's options extracts from HTML you send instead of downloading the page, as on Zyte (the url is still required):

{ "url": "https://shop.example/item", "product": true,
  "productOptions": { "extractFrom": "userHtml", "userHtml": "<html>…</html>" } }

When every output is such an extraction and runs locally (the built-in markdown, or your extractors), there's no request at all. With zyteApi() and no local extractor, Zyte does it.

Models

openAICompatible() speaks the Chat Completions API with response_format: json_schema:

Provider	baseURL
OpenAI	default
OpenRouter	https://openrouter.ai/api/v1
Groq	https://api.groq.com/openai/v1
Together	https://api.together.xyz/v1
Anthropic (OpenAI-compatible endpoint)	https://api.anthropic.com/v1/ (may ignore response_format; the prompt and parser still get JSON)
Ollama / LM Studio / vLLM (local)	http://localhost:11434/v1 …

For servers without JSON-schema output, pass responseFormat: 'json_object'. Any other SDK works through complete directly:

llmExtractors({
  complete: async ({ system, prompt, schema, field }) => {
    const text = await myModel.generate({ system, prompt, jsonSchema: schema });
    return { output: text, usage: { inputTokens: 0, outputTokens: 0 } }; // or just return the text / object
  },
  fields: ['product', 'article'], // only these; others stay 400
});

From env, for xyte serve: XYTE_LLM_MODEL, XYTE_LLM_BASE_URL, XYTE_LLM_API_KEY, XYTE_LLM_RESPONSE_FORMAT, XYTE_LLM_MAX_INPUT_CHARS.

Compared with Zyte

Zyte runs purpose-trained models and also uses visual features of the rendered page. A general-purpose LLM over Markdown gets the common fields (name, price, currency, images, description, author, dates) right on most pages, but quality depends on the model, and every field costs a model call. markdown itself needs no model: it's built in.