Cloudflare releases Clairic, a System One extraction model for structured data

What happened

Cloudflare announced a preview of Clairic, a new class of model called a System One extraction model. The model is available now on Workers AI under the handle @cf/cloudflare/clairic, with an open-weights release planned for the future. Clairic complements existing System One decision models (Jev and Clef, made by TypeSafe) by addressing a different set of tasks: extracting, computing, estimating, and forecasting values from unstructured inputs like text and images.

How System One extraction models differ from LLMs

System One extraction models operate differently from large language models. While LLMs generate text token-by-token through multi-step reasoning, Clairic processes input in a single forward pass and returns typed values directly. This single-pass design eliminates the need to charge for output tokens. Clairic also cannot hallucinate in the same way as an LLM because it returns a value of the declared type or null, not free-form text that requires parsing and validation.

Traditional LLMs face several failure modes when asked to extract structured data: JSON output may be malformed, generated digits can be misread, and there is no way to trace extracted text back to its source location. By contrast, Clairic returns character offsets showing exactly where in the input each extracted value came from, confidence scores to indicate which answers warrant a second look, and strongly typed outputs where numbers and dates follow their type rules (for example, currency amounts with two decimal places).

What the model can extract and compute

Clairic returns four kinds of values. Extracted values are text spans copied directly from the input, with their character offsets and confidence scores. Computed values are numbers, dates, times, datetimes, durations, and lists that the model reads from the input or calculates from it. Estimated values are numbers the input doesn't explicitly state but gives evidence for; these return as distributions with a median and an 80% confidence interval. Forecasted values predict future points in a series, each with its own 80% interval.

A request consists of unstructured text plus up to four optional images and a set of typed fields. Each field has a name, a type, and optional plain-language instructions. Clairic processes all fields in parallel in a single pass and returns one typed answer per field. There is no limit on the number of fields; if they do not fit in one pass alongside the input, the server splits them into parallel passes that each see the whole input, then merges the results. Every scalar answer includes a probability that the value is absent (p_null) and a confidence score. Estimate answers include a median, an 80% interval (p10 to p90), and all 99 percentiles on request.

Practical examples of extraction

The source includes an example where Clairic processes a customer support email asking for a refund. The email contains an order ID, delivery date, amount paid (which the model computes by adding price and tax), number of days since delivery (computed from two dates), and a list of damage types. Clairic extracts the order ID and damage list with character offsets, computes the total amount and days elapsed, and returns null for a tracking number that does not appear in the email. Each answer includes a confidence score showing how sure the model is; for the tracking number, the model shows a guess of what it would have returned if forced to choose, but correctly identifies null as the answer.

Clairic also processes images. In an example with a photograph of a receipt, the model extracted the order number, purchase date and time, quantity, unit price, subtotal, sales tax, and tip, all with confidence scores. It correctly identified that no discount was present and returned null for that field. This demonstrates that the model can handle imperfect image quality and angles while still returning typed values.

Performance and cost implications

Because Clairic operates in a single pass and charges only for input tokens, it is significantly faster and cheaper than using an LLM for the same extraction tasks. Traditional LLM-based extraction pipelines typically include multiple layers of parsing, schema validation, retries, and spot checks to combat failure modes like malformed JSON or misread digits. Developers pay for reasoning and output tokens on every field of every document, which adds up quickly when processing many documents or requesting many fields per document. With Clairic, requesting thirty fields costs approximately the same as requesting three, since all fields are answered in the same parallel pass.

Extraction is one of the most common tasks sent to frontier LLMs today, according to the source. Common questions include invoice totals, delivery dates, lateness calculations, repair cost estimates, and traffic forecasts. None of these have a fixed set of answers that could be listed in advance, which makes them unsuitable for decision models that return probabilities over predefined options. This category of open-valued extraction and estimation is where Clairic fills a gap between decision models and general-purpose LLMs.

Source

AI-assisted, verified against the source. This post was drafted with AI from the primary source linked above and checked automatically against it. For anything you depend on, read the original.