Cloudflare Launches Clairic to Extract Data Without a Chatbot

Cloudflare has launched a preview of Clairic, a specialist AI model for pulling structured facts and numerical estimates out of messy text and images. Available on Workers AI, it is built for tasks such as finding an invoice total, identifying a date in an email or estimating a value from the evidence in a document. The pitch is that many businesses do not need a chatbot to write a paragraph when their software needs a dependable field to put into a database.

In its October 19 announcement, Cloudflare calls Clairic an extraction model in a new System One class. Unlike a general-purpose large language model, it returns typed values according to a schema supplied by the developer. Those values can include copied text spans, numbers, dates, lists, estimates and simple forecasts. Cloudflare says the model also provides confidence information and can return null when it cannot find a requested value.

A conventional language model generates its response token by token. Clairic is designed to process a request in a single pass and answer several structured fields together. Cloudflare argues this can reduce latency and cost in repetitive extraction workflows, though its comparisons are company-run and may not predict every customer’s results. The preview is hosted as @cf/cloudflare/clairic on Workers AI; Cloudflare says it intends to release the weights openly later, but they are not available yet.

For a developer, the useful distinction is between extracting a value and asking a model to explain it. An accounts team may need an invoice number, amount and due date. A logistics system may need delivery times from unstructured messages. A support workflow may need a product identifier or a list of items. Clairic is intended to return those fields in defined types so the next step of the application does not have to parse free-form prose.

Cloudflare says the model can handle text and up to four images in a request, which makes photographed receipts one possible input. It also returns source offsets for extracted text, allowing software to trace a value back to its place in the original input. Confidence scores are useful for deciding when a human should review a result, but they should not be treated as proof that an answer is correct.

Cloudflare is unusually direct about the preview’s limitations. It says multi-step arithmetic is not yet strong, estimates need evidence, calibration is not guaranteed on unfamiliar data and forecasting is basic. Its advice is to let code perform complex calculations and to use the model’s confidence signals to decide which answers require review. That is especially important if extracted numbers feed payments, medical records or other consequential decisions.

Clairic also extends a broader Cloudflare strategy. Earlier this month, TechBooky covered its Clef models for fast, structured decisions. Clef chooses among defined options; Clairic aims to supply the values those decisions may depend on. General-purpose models can still handle writing and open-ended reasoning. In a practical AI workflow, the three approaches may work together instead of one replacing the others.

The commercial question now is whether Clairic performs reliably on documents outside Cloudflare’s demonstrations. Developers will need to test accuracy, failure handling, latency and real operating cost against existing extraction pipelines. For businesses processing large volumes of routine documents, a focused model could be valuable. The preview is a starting point, not proof that every extraction job should move away from a large language model.