Cloudflare unveiled a preview version of its AI model, Cleric, that can extract information from documents and images quickly and at low cost, it said on its official blog on Oct. 19.

The company said Cleric can be used on Cloudflare's AI platform, Workers AI. Cloudflare plans to soon release it in an open-weights form that anyone can download and use.

Cleric is a "System 1"-type model. System 1 is a concept drawn from Daniel Kahneman's book "Thinking, Fast and Slow" and refers to models that make quick and accurate judgments. Large language models, or LLMs, that reason deeply and write text correspond to "System 2."

Cleric extracts names, dates and amounts from unstructured materials such as emails and photos of receipts. It also calculates or estimates values that are not in the material and predicts figures that may come next.

It does not generate text one character at a time like an LLM and instead produces an answer in one shot. Charges apply only to input data, and most answers come within 1 second. If it cannot find an answer, it does not fabricate a plausible value and marks it as "none." It also provides a confidence score with each answer.

For example, if a customer complaint email lists a product price of $649 and tax of $52.1, Cleric adds them and returns a total payment amount of $701.1. It also calculates elapsed days by comparing a delivery date with the email's 작성 date.

Cloudflare tested a method in which a decision model, Cleft, takes the values extracted by Cleric to make a determination, and an LLM writes the reply. The company said the method was 3.2 times faster and cost 49 percent less than a method that left every step to the LLM GLM-5.3.

Cloudflare acknowledged, however, that it is still weak at calculations that involve multiple steps. It added that it cannot properly estimate values without a basis and that its time-series forecasting is at a basic level.