Typed Extraction Model Clairic Now Available on Workers AI

Workers AI now includes Clairic, a preview‑only extraction model that returns typed values rather than raw text. For AI engineers, platform teams, and SREs, this means you can ask the model to pull exact numbers, dates, or estimates from a document and receive them in a predictable schema, cutting the need for custom parsers and post‑processing logic.

How the Typed Extraction Model Works

Clairic expects two inputs: a state string containing the raw text (or up to four images) and a fields object that defines the desired output types. Each field specifies a type such as text, number, date, duration, list, estimate, or forecast. The model processes the input in a single pass and returns one value per field. If the input does not determine a value, the field is set to null. Every scalar answer includes a p_null probability and a confidence score; estimates and forecasts also provide a median and an 80 % interval (p10–p90), with an optional flag to include full quantiles.

Practical Implications for Architecture and Operations

Integration is a single call to the Workers AI runtime. The example below shows a typical JavaScript invocation inside a Cloudflare Worker:

const response = await env.AI.run("@cf/cloudflare/clairic", {
  model: "clairic",
  state: "Invoice INV-2291 from Northwind Supply, issued 2026-09-02, net 30. 4 pallets at $250.00 each, shipping $80.00. Contact: Dana Ruiz, Priya Shah.",
  fields: {
    invoice_number: { type: "text" },
    issued: { type: "date" },
    due_date: { type: "date", instructions: "Issue date plus the net payment days" }
  }
});

Because the model returns typed data, downstream services can skip regex‑based extraction and directly consume the values. Parallel passes are used automatically when the schema exceeds what can be read in a single pass, so there is no hard limit on the number of fields. The ability to attach up to four images expands use cases to scanned documents or screenshots without additional preprocessing.

Security and Governance Considerations

Clairic does not generate free‑form tokens; it only emits values that conform to the declared type or null. This deterministic output reduces the attack surface associated with uncontrolled text generation. However, the model still returns confidence metrics, which should be treated as part of a risk‑based decision process. Practitioners need to define thresholds for automation versus manual review, especially when p_null or low confidence scores appear. Image inputs are limited to four per request, so any workflow that relies on visual data must respect that bound and consider the privacy implications of sending images to the model.

Related CloudNinjas coverage: AI engineering.

What This Means For Practitioners

• Identify pipelines that currently parse invoices, logs, or reports for numeric or date fields and replace custom parsers with a Clairic call.
• Prototype the model in a non‑production Worker to observe confidence and p_null distributions for your data.
• Establish automation thresholds based on the returned confidence scores; route low‑confidence results to a review queue.
• Plan for handling null values gracefully in downstream services.
• Monitor the upcoming open‑weights release, which may affect model behavior or licensing.