Cloudflare has introduced Clairic, a preview AI model designed to turn unstructured information into typed values such as names, dates, numbers, estimates and forecasts.
The model is available through Workers AI under the identifier @cf/cloudflare/clairic. Cloudflare describes it as the first of a new class of “System One” extraction models: narrowly focused systems intended to answer in a single pass instead of producing a token-by-token explanation.
From documents to typed values
Clairic accepts raw text together with a developer-defined schema describing the required fields. Its responses can include copied text spans, calculated dates and durations, lists, bounded estimates and time-series forecasts. Extracted fields include character offsets and confidence scores, while estimates and forecasts include a median and an 80% interval.
That structure is meant to reduce the parsing and validation work required when a general-purpose large language model returns free-form text. Clairic must return a value matching the requested type or return null when it cannot find an answer. Cloudflare cautions that the model can still make mistakes, so confidence signals and application-level checks remain important.
A different role from an LLM
Cloudflare positions extraction models alongside, rather than as replacements for, LLMs. One model can extract facts from an email or document, another can select an action from a fixed set, and an LLM can produce the final explanation or response.
The company says Clairic typically responds in under a second and charges only for input because it does not generate output tokens in the conventional way. That could make it useful for high-volume workflows involving invoices, claims, order records, service tickets or operational forecasts.
In one example described by Cloudflare, the model reads a damaged-product claim and returns fields including an order number, amount paid, delivery interval and estimated repair cost. A separate decision model can then apply company policy before an LLM writes a customer-facing message.
Preview status matters
Clairic is currently a hosted preview on Workers AI, not a generally available service with a long production history. Cloudflare says it intends to release the model as open weights, but it has not provided a release date.
Teams evaluating the preview should test it against their own document formats, languages and error costs. A fast typed response may simplify an automated pipeline, but confidence thresholds, fallbacks and human review are still necessary when a wrong number or date could trigger a consequential action.
Source: Cloudflare: Introducing Clairic, a new System One class of extraction models.
Featured image: AI-generated editorial illustration by Bex News.
