Cloudflare Clairic: Structured Data Extraction Cuts LLM Costs by 49% and Latency

Cloudflare released a preview of Clairic, an extraction model that pulls typed values from messy text and images in a single pass without generating tokens. This tool targets routine developer tasks, such as reading invoices to return order numbers, totals and due dates as JSON, which are often assigned to large language models.

Root cause

Developers frequently use large language models to extract specific fields from unstructured data, but these models can emit broken JSON, misread digits or return excerpts without source pointers. Teams often stack parsers, schema checks and retries on top of these calls, paying output tokens for every field extracted, which increases both cost and complexity.

Attack path

Clairic runs on Workers AI and accepts text, up to four images and a list of named fields with defined types such as number, date or estimate. Answers arrive as a value of that type or as null, with scalar answers carrying a confidence score and the probability that the value is missing.

Affected versions

The model is available as @cf/cloudflare/clairic on Workers AI. Cloudflare plans to publish the weights for this preview release. The system charges only for input processing, rather than for the output tokens that large language models typically require for each extracted field.

Mitigation

• Review current extraction pipelines to identify where large language models are used for simple field extraction.
• Test Clairic on sample invoices or support claims to compare latency and cost against existing methods.
• Verify that extracted values include character offsets and confidence scores for validation purposes.
• Monitor for updates on the publication of model weights if open-source inspection is required.

Why it matters

Teams often rely on large language models for data extraction because they are versatile, but this versatility comes with hidden costs and errors. Understanding the difference between generative models and dedicated extraction tools helps engineers choose the right instrument. Generative models predict the next token, which can introduce hallucinations in structured data. Extraction models focus on locating existing information, reducing the risk of fabricated values. Checking whether a task requires generation or simple retrieval can significantly lower operational expenses and improve accuracy.

What to do and how to stay safe: Cloudflare

• Audit current workflows to identify tasks that involve extracting structured data from unstructured sources.
• Evaluate whether large language models are being used for simple extraction tasks that could be handled by specialised tools.
• Monitor vendor announcements for new extraction models that offer lower latency and cost structures.
• Once the vendor provides an update, test the new model against existing pipelines to measure performance improvements.

Step-by-step guide: Stop Sensitive Data Leaks in AI Chatbots: Priority Controls

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

Frequently asked questions

How does Clairic differ from large language models in data extraction?

Clairic pulls typed values in one pass without generating tokens, whereas large language models generate output tokens for each field, increasing cost and potential for error.

What are the performance benefits of using Clairic?

In Cloudflare’s test, a pipeline using Clairic finished in 19.3 seconds compared to 61.7 seconds for an LLM version, and cost 49% less.

Does Clairic provide confidence metrics for extracted data?

Yes, scalar answers carry a confidence score and p_null, the probability that the value is missing, while extracted text includes character offsets.