Cloudflare previews Clairic to extract numbers, dates and lists from input
Cloudflare has put its Clairic extraction model into preview on Workers AI. Developers supply input and define the fields and value types they need; the model returns values instead of a free-form explanation. Null responses, confidence scores and uncertainty intervals help separate automated results from those requiring review. Cloudflare identifies limits in long calculations and table aggregation. It plans an open-weights release at a later stage.
Artificial Intelligence··Morning
Clairic returns a value for each requested field
Cloudflare, a provider of internet infrastructure and developer services, put Clairic into preview on its Workers AI model-hosting platform on October 19.[1], [2]
The model is designed to extract values from supplied input. Developers send a state and a schema specifying the fields they need and the type of value each field accepts. Clairic returns values for those fields in place of a free-form explanation. Cloudflare’s earlier decision models select among options listed in advance; this release addresses fields such as an invoice total or an unknown delivery date whose possible answers cannot necessarily be enumerated beforehand.[1]
Types determine how answers are returned
Supported values include text, numbers, dates, times, durations and lists. A text answer returns an exact span from the input together with its character positions. Numeric fields can impose integer or decimal constraints, and a list can preserve the order of matching elements in the original input. Estimate fields produce a distribution inside limits chosen by the developer. Instead of generating output tokens one after another, Clairic reads the input and decodes answers using the rules of their types. Where the input cannot determine an answer, the model can return null, an empty value, rather than filling the field with a result. Combined date-and-time fields are also supported. Input can include up to four images, and time-series forecasts can extend up to 64 future steps. Developers determine the lower and upper bounds for estimates, which are returned as distributions within those limits.[1]
Long calculations remain a task for code
Confidence values, the probability of null and uncertainty intervals are provided to help developers decide which results to automate and which to send for human review. Correct formatting remains separate from a correct answer. Cloudflare identifies weaknesses in long chains of dependent calculations and in aggregation across many table rows. For these tasks it recommends extracting the inputs and performing the mathematics in code. The company also says longer forecasts can flatten toward the average, miss holidays absent from the series and produce intervals that are too narrow on unfamiliar data. The current release is a hosted preview; Cloudflare describes publishing the model weights as a later intention.[1]
