Run Cloudflare's Clairic on Workers AI to slash extraction costs by 49%

Cloudflare has released a preview of Clairic, a specialized model designed to pull specific data points from messy text and images without generating conversational output tokens. According to Help Net Security, the company aims to reduce the high costs and latency associated with using general-purpose large language models for routine extraction tasks. This tool targets developers who frequently ask AI to parse invoices or support tickets for structured information like dates, totals, and order numbers.

In plain English

Large language models are often used to read documents and return specific facts, but they can be expensive and slow for this job. They generate text token by token, which means you pay for every word they write, even if you only need a single number. Clairic is different because it does not generate free-form text. Instead, it looks at the input and returns only the requested values in a predefined format. This approach avoids the overhead of generating unnecessary output, which helps lower costs and speed up the processing time for simple data retrieval tasks.

The background

In a test involving a damaged-product support claim, Cloudflare compared a pipeline using Clairic against one that relied entirely on a large language model. According to Help Net Security, the Clairic-based pipeline finished in 19.3 seconds, while the LLM-only version took 61.7 seconds. The cost difference was also significant, with the new method costing 49% less. The test showed that Clairic extracted facts in 0.46 seconds, and another model, Clef, made a decision in 0.22 seconds, whereas the comparable LLM steps took much longer.

Clairic runs on Workers AI and allows developers to send text and up to four images along with a list of named fields. Each field has a type, such as a number, date, or list. The model returns a value of that type or null if the data is missing. Scalar answers include a confidence score and a probability that the value is absent. Cloudflare charges only for the input tokens, not for the extracted values. The company plans to publish the model weights in the future.

What changes now

Developers can now access Clairic through the @cf/cloudflare/clairic identifier on Workers AI. This preview allows teams to test the model on their own workflows, such as processing customer emails or invoices. In Cloudflare’s sample tests, Clairic correctly calculated totals not explicitly stated in emails and identified missing tracking numbers by returning null. The model also demonstrated consistency in estimation tasks where other models varied their answers. Teams can evaluate whether switching from general LLMs to this specialized extractor improves their performance and budget.

Key terms explained

Structured data extraction involves pulling specific information from unstructured sources like emails or images. Teams usually define the fields they need, such as dates or prices, before processing. Confidence scores help administrators decide if a result is reliable enough to use without human review. Null values indicate the model could not find the requested data, which is often more useful than a guessed answer. Checking for these indicators helps prevent errors in downstream systems that rely on accurate input.

What to do and how to stay safe: Cloudflare

• Review current workflows that use large language models for simple data extraction tasks to identify potential cost savings.
• Test the new extraction model on a small batch of internal documents to compare speed and accuracy against existing solutions.
• Monitor confidence scores and null values returned by the model to determine if human verification is still required for critical data.
• Wait for the vendor to publish model weights before making long-term infrastructure commitments that depend on open access.

Step-by-step guide: Stop AI Chatbots Leaking Sensitive Business Data

General security guidance from the Vector Update newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

How does Clairic differ from standard large language models?

Clairic is designed specifically for extraction and does not generate output tokens, whereas standard LLMs generate text token by token, which increases cost and latency.

What inputs can Clairic process?

Developers can send text and up to four images to the model, along with a list of named fields with specific types like number, date, or list.

How much does Cloudflare charge for using Clairic?

Cloudflare charges only for the input tokens, not for the extracted values or output, according to the preview details.