Overview

Cloudflare has unveiled Clairic, a new extraction model available on Workers AI as @cf/cloudflare/clairic. It is described as a System One extraction tool that pulls, computes, estimates, and forecasts numeric or typed values from unstructured input. The model focuses on single‑pass inference without the token‑by‑token generation typical of large language models.

Clairic processes plain text, PDFs, or JSON blobs and returns outputs that are tightly typed. A user can ask for a number, such as an invoice total, and Clairic will extract the numeric value, or compute a derived figure, returning it with two decimal places and an 80 % confidence interval. If the input contains a name or an order ID, Clairic outputs a span of characters, including the character offsets and a confidence score. Dates, times, datetimes, durations, and lists can be extracted in the same way, each with type‑specific formatting rules. When the model cannot find a value, it responds with null; otherwise, it includes a confidence score for each result, helping users decide whether to trust the answer or review it manually. For estimations, a user declares an upper and lower bound and a grid; Clairic produces a distribution over that range with a median and an 80 % interval. Forecasting works similarly, returning a single estimate per step ahead, each with its own interval.

Typical inputs include email threads about order status, invoices, or product descriptions. Outputs are numeric values, date strings, or spans of extracted text. The model is well suited for developers and data pipelines that need fast, precise structured outputs from noisy documents, such as e‑commerce order processing or logistics tracking. It is not intended for tasks that require free‑form language generation, deep reasoning, or elaborate text transformation. The listing does not provide any pricing information, deployment options beyond Workers AI, or details on training data or privacy safeguards.

Key features

Sub‑second, single‑pass inference with lower cost than token‑based LLMs
Typed output: spans, numbers, dates, times, lists, and probability distributions
Confidence scores attached to every extracted, computed, estimated, or forecasted value
No token generation, reducing hallucination risk
Hosted on Workers AI as @cf/cloudflare/clairic; open‑weights announced for future release