Cloudflare previews Clairic, a model for extracting and estimating typed data

Cloudflare has introduced Clairic, a preview model on its Workers AI platform designed to pull structured information from unstructured text and images. The company says the model can extract or calculate values, estimate information that is not stated directly, and forecast future values. Cloudflare says it intends to release the model’s weights but did not give a date in its announcement.

Clairic belongs to what Cloudflare calls “System One” models: systems built for fast, narrowly defined tasks, as opposed to general-purpose language models that generate text and can reason through a prompt. Clairic returns typed outputs such as names, dates, numbers and lists, while Cloudflare’s existing decision models, Jev and Clef, choose among options and supply probabilities.

A Clairic request pairs source material with a set of fields and instructions. Cloudflare says the model processes those fields in a single forward pass and returns values in the requested formats, along with confidence information and, for scalar answers, an estimate of the likelihood that the information is absent. For extracted text, it can also return character offsets pointing back to the source. When it cannot find a value, it can return null rather than a generated answer.

The approach is aimed at routine data-processing work such as reading invoice totals, delivery dates or support requests. In an example published by Cloudflare, Clairic adds a listed price and tax to calculate a total, works out the number of days since delivery from two dates, and marks a missing tracking number as null. These demonstrations illustrate the advertised workflow; they are not an independent evaluation of the model’s accuracy.

Cloudflare says Clairic charges for input rather than generated output tokens and typically responds in under a second. It also claims the model can be faster, cheaper and more accurate than traditional large language models for suitable extraction tasks. The announcement does not include comparative benchmark results to support those claims, and the model remains in preview.