Clairic

The autoregressive generation characteristic of LLMs determines that their output is streamed token by token, stemming from the decoding strategy of the Transformer architecture.

Cloudflare Workers AI has multiple embedding models built in, with all model inference completed by Cloudflare Workers AI.

Clairic is priced at $0.10 per million input tokens, with no charge for output.

For Clairic's list answers, values are returned with spans and confidence scores; estimated values return median, p10, and p90 instead of p_null and need to be checked separately.

Clairic is provided "AS IS" and "AS AVAILABLE," and Cloudflare may discontinue the service at any time without refund; the weights are not yet public, and it currently runs only on Cloudflare endpoints.

In Clairic's scalar answer object, the value field is the actual answer, p_null represents the probability that the input does not contain that value, and guess is not the answer; when value is null, guess should not be read.

In the official example, the answer object for po_number contains type: text, p_null: 1, span: [8, 16], guess: INV-2291, confidence: 0.9995, value: null.

Cloudflare's official documentation states that a null single-value result means the model considers the input information insufficient.

Clairic is not good at multi-step arithmetic, and Cloudflare recommends having Clairic predict values and letting code do the math.

In the official example response, 61 state_tokens are used for the invoice text and 472 schema_tokens for the field definitions, totaling 655 input_tokens, with output_tokens of 0 and passes of 3.

Clairic is a zero-output-token extraction model that returns one value per field in a single forward pass and does not generate text.

Clairic handles 8,192 tokens of fields plus input per call, with additional fields split into parallel batches; it supports up to 4 embedded images (PNG, JPEG, or WebP) and does not support remote URLs.

Clairic handles four types of tasks: extracting values already present in the input, computing values from facts in the input, estimating values the input does not pin down, and predicting sequences.

Clairic positions itself as a "System One extraction model" that returns typed values with confidence scores in a single forward pass.

Clairic's "single forward pass" means that in a single call the model, through multiple output heads or structured prediction heads, simultaneously gives the values and confidence of all target fields without token-by-token iteration.

Clairic is currently in preview, and the calibration quality of its confidence scores, the coverage of typed values, and its cross-domain generalization still require actual testing to judge.

For latency-sensitive, high-volume extraction scenarios, Clairic is worth adding to the evaluation list, but a more realistic positioning is as the first filter in the pipeline rather than a complete replacement for general-purpose models.

System One is defined by Clairic as an entirely new model category focused on extracting information from data.

Clairic posted on Hacker News its extraction model, which it calls the "System One" category.