Clairic by Cloudflare: Data Extraction from Documents 49% Cheaper than LLMs

Cloudflare Reduces Data Extraction Costs from Documents by 49% with Clairic

Cloudflare has introduced Clairic, an extraction model that extracts structured values from messy documents in a single operation, without generating tokens. In internal tests, Clairic processed a support request for damaged products in 19.3 seconds, compared to 61.7 seconds for an LLM-based system, saving 49% on costs.

The traditional system using LLM to extract data from documents like invoices or emails presents numerous problems. These models can generate invalid JSON, misread figures, or return extracts without source references. To mitigate these errors, companies must integrate additional parsers, schema checks, and repeated attempts, increasing operational costs.

Quick Answer

How Clairic Works

Clairic operates on Workers AI and can process text, up to four images, and a list of named fields with their types (number, date, list, or estimate). Results are returned as values of the requested type or null. Scalar values include a confidence score and pnull, the probability that the value is missing from the input. Extracted text is accompanied by character offsets. Cloudflare charges only for the processed input.

In a test on a customer email, Clairic calculated a total of $701.10, summing a price of $649 and a tax of $52.10, even though the total was not explicitly stated in the email. It counted 11 days between delivery and the email date. For a tracking number not present in the email, it returned null, with a pnull value of 1.

Comparison with LLM-Based Solutions

In the refund request processing test, Clairic extracted facts in 0.46 seconds, while Clef, Cloudflare's model for option selection, chose between repair or refund in 0.22 seconds. GLM-5.3, the LLM used in the comparison, took 38 seconds for the same operations. Both solutions used GLM-5.3 to generate the response, which took 18.6 seconds of Clairic's total workflow processing time.

The Clairic-based workflow cost $6,032 per million requests, compared to $11,740 for the LLM-based system. Both systems provided correct results in every execution reported by Cloudflare. Clairic maintained a consistent repair estimate, while GLM-5.3 returned different values in repeated executions. Gemma 4 and Qwen, used as LLM-based pipelines, failed to determine the repair cost.

Expected Applications and Limitations

Cloudflare has also tested Clairic for predictions, such as counting TSA checkpoints in U.S. airports. Over 21 days of predictions, Clairic had an average error of 3.3%, with 17 of the 21 actual counts within the 80% confidence interval. Three of the four errors occurred during the Columbus Day weekend, an event the model could not predict based on data history.

Cloudflare emphasizes that Clairic is still in development and has some limitations. Complex mathematical operations or sums across multiple rows of tables may not be reliable, and the company advises using specific code for these operations. Additionally, prediction intervals may be narrow on unfamiliar data, and predictions cannot replace specific time series models for the data in question.

Implications for Security and Data Management

Clairic's ability to extract structured data from messy documents can have significant implications for security and data management. For example, integrating Clairic into a corporate DLP system could enhance the ability to detect and protect sensitive information within unstructured documents.

For companies managing large volumes of documents, adopting Clairic could significantly reduce operational costs related to data extraction and validation. This could be particularly useful for incident response teams, which often need to quickly analyze large amounts of information to effectively respond to security threats.

Furthermore, using Clairic could improve the efficiency of compliance processes, reducing the time and resources needed to extract and verify data required by regulations.