Extract API v2: schema-guided agents with self-healing selectors.

Every extraction pipeline we've ever seen dies the same way. It works on Tuesday. On Wednesday the site ships a redesign, a class name changes from price to price-v2, and the data silently goes empty. Someone notices on Friday.

Extract API v1 helped by moving parsing into a model instead of selectors. v2 goes further: extraction now plans its own path through a page, verifies what it found and remembers how it found it.

What changed

• Planning. Given a schema, the model decides whether to scroll, paginate or click "load more" before extracting.
• Self-healing selectors. We cache a fast selector for each field and re-derive it automatically when it stops matching.
• Confidence and provenance. Every field returns a confidence score and the DOM snippet it came from.

const result = await tessera.extract(page, {
  schema: z.array(Product),
  plan: "auto",          // scroll, paginate, expand
  minConfidence: 0.9,    // flag anything below
});

result.data;        // Product[]
result.confidence;  // per-field scores
result.selectors;   // cached for next run

How self-healing works

On the first run, the model reads the rendered page and returns data plus a minimal CSS or XPath selector for each field. On later runs we try the cached selectors first, which is fast and cheap. If a selector returns nothing, or the value fails validation, we fall back to the model, re-derive the selector and update the cache.

“We went four months without touching a selector. The pipeline just kept healing itself.”

In our benchmark of 1,200 retail and travel sites over 90 days, v2 kept 99.2% of fields flowing through redesigns that broke every hand-written scraper in the control group.

Pricing and availability

Extract API v2 is live today for every plan. Cached-selector runs cost 80% less than full model extractions, so most pipelines get cheaper automatically. Existing v1 calls keep working, and you can opt in with version: 2 or by upgrading the SDK.