Inspired by TypeSafe AI’s System One Model Jev, Cybersecurity firm Cloudflare introduced Clairic, a new AI model designed to extract, calculate, estimate and forecast values from unstructured data.
Cloudflare has launched Clairic under preview version on Workers AI and claims they will release its full weights soon. According to Cloudflare, Clairic delivers higher accuracy, faster responses and lower costs than traditional large language models (LLMs) for specific tasks.
Cloudflare claims that unlike System One models like Jev or its in-house Clef AI model, Clairic is first of its kind system one extraction model that can return extracted, compute estimate and forecast values from text, numbers, dates, times, datetimes, durations and lists, copied from the input.
Cloudflare introduces a new workflow where System one model Clairic performs extraction role by turning raw input into typed values, combined with another System one model Jev or Cloudflare’s in-house Clef model for turning values into predictions and then using traditional LLMs like Claude or ChatGPT for turning decisions into words.
Cloudflare stated that they got inspired by the recently introduced Jev System One Models, that do not produce free-form texts like LLMs but returns structured answers with probabilities and confidence scores, allowing software to use its decisions directly.
“The introduction of Jev prompted people to start thinking about a new class of models that were not LLMs and how to integrate them into their workflows. We were inspired by the concept of System One models and started thinking about other tasks where LLMs are inefficient,” read an excerpt from a blogpost by Cloudflare, explaining the genesis of Clairic extraction model.
System one model is an AI model or an operating system that aims at producing fast answers with little reasoning. This term is found in theories of two types of human thinking, written by Daniel Kahneman’s, which are called System 1 and System 2. System 1 thinking is fast, automatic, and intuitive, whereas System 2 involves slow and more thoughtful consideration.
According to Cloudflare the model can extract information such as names, order IDs and dates from text, calculate values such as invoice totals and delivery delays, estimate costs based on available information, and forecast future values in a time series.
Unlike conventional LLMs, Clairic returns structured, typed values rather than generated text. Cloudflare said Clairic processes requests in a single forward pass without generating output tokens. Users can submit text, up to four optional images and a set of fields specifying the values they want to retrieve.
The model returns answers for those fields in parallel, with the server splitting requests into additional parallel passes when necessary.
Cloudflare positioned Clairic alongside its existing system one decision models, including Clef, which selects from a fixed set of options and assigns probabilities to them. The company acknowledged that Clairic can still make mistakes, despite not generating tokens in the same way as an LLM. Its confidence scores are intended to help users identify outputs that may require further review.
Also Read: Jev AI Reaches 35% of Paid Teams in 14 Days, Uses Just 0.15% of Inference Spend: Vercel Report
