Cloudflare Clairic

Product Introduction

1. Definition: Cloudflare Clairic is a System One extraction model, a new class of AI model designed for structured data extraction from unstructured text and images. It is a regression and extraction model that performs a single forward pass on input tokens, returning typed values without generating any output tokens.
2. Core Value Proposition: Clairic exists to provide developers with a faster, cheaper, and more accurate alternative to using large language models (LLMs) for structured data extraction tasks. It directly addresses the inefficiency, cost, and hallucination risks of using generative AI for tasks like invoice processing, data entry, and forecasting.

Main Features

1. Structured Value Extraction: Clairic extracts strongly typed values directly from input text or images. It returns data in predefined types like text, number, date, time, datetime, duration, and list. Each extracted value includes character offset spans for traceability and a calibrated confidence score. This feature works by analyzing the input in a single parallel pass, grounding every answer in the source material.
2. Computation and Estimation: Beyond simple extraction, Clairic can compute values not explicitly stated. For example, it can sum numbers mentioned separately (e.g., price + tax) or calculate durations between dates. For predictive tasks, it offers an estimate type that returns a probability distribution (median, p10, p90) within user-defined bounds, providing a calibrated view of uncertainty for values like repair costs or property prices.
3. Time Series Forecasting: The forecast field type allows Clairic to predict future values in a series, such as daily sales or website traffic. Given historical data in plain text or CSV format, it returns a forecasted value and an 80% prediction interval for each future step (e.g., the next day). This is implemented as a specialized forward pass over the sequential data, offering a middle ground between simple heuristics and complex, dedicated time-series models.

Problems Solved

1. Pain Point: The high cost, latency, and unreliability of using frontier LLMs for repetitive structured data extraction. LLMs generate tokens, which is expensive for simple lookups, and can hallucinate or output malformed JSON, requiring extensive validation pipelines.
2. Target Audience: Developers and engineers building data processing pipelines, especially in fintech, logistics, customer support, and business intelligence. Specific personas include Backend Engineers automating document processing, Data Scientists needing quick feature extraction from text, and Product Managers implementing AI features without LLM overhead.
3. Use Cases: Essential for processing invoices and receipts (extracting totals, dates, line items), parsing customer support emails (identifying order IDs, issues, dates), estimating values from textual evidence (e.g., property value from comparable listings), and generating short-term forecasts from historical data logs.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):

   • Pros: Extremely low latency (sub-second responses) and cost ($0.10 per million input tokens, no output token cost). High accuracy for extraction tasks with calibrated confidence scores. Eliminates JSON parsing and schema validation issues by returning native typed values. Cannot hallucinate text in the traditional sense, as it doesn't generate tokens.
   • Cons: Limited to structured output; cannot generate open-ended text or perform complex reasoning. Forecasting capability is general-purpose and may not match the accuracy of a finely-tuned, domain-specific time-series model. Requires a predefined schema (fields), so it's not suitable for exploratory or ad-hoc queries.

2. Key Alternatives & Differentiation:

   • OpenAI GPT-4 / Claude for Extraction: These LLMs are flexible but slower, more expensive (charge for input and output tokens), and prone to formatting errors. Clairic is purpose-built, cheaper, faster, and guarantees typed output.
   • Traditional OCR + Rule-Based Pipelines: Solutions like AWS Textract or Google Document AI extract text well but require custom code for computation, estimation, and entity linking. Clairic unifies extraction, computation, and estimation in a single API call.
   • Other System One Models (Jev, Clef): These are decision models for classification/ranking. Clairic is the first System One extraction model, solving a complementary problem—turning unstructured data into typed values—that feeds directly into decision models.

Frequently Asked Questions (FAQ)

1. How does Cloudflare Clairic's pricing compare to using GPT-4 for data extraction? Clairic charges $0.10 per million input tokens with no charge for output, making it significantly cheaper than GPT-4, which charges for both input and generated output tokens, especially for tasks requiring structured JSON responses.
2. Can Cloudflare Clairic process handwritten documents or scanned PDFs? Yes, Clairic accepts images as input (up to four per request). It uses integrated vision capabilities to extract text and data directly from photos of documents, like receipts or forms, without a separate OCR preprocessing step.
3. What is the difference between an "extraction model" and a "decision model" like Clef? An extraction model (Clairic) answers "what is the value?" by pulling or computing numbers, dates, and text from data. A decision model (Clef) answers "which option?" by choosing from a predefined list. They are designed to work together in a pipeline: Clairic extracts data, and Clef makes a decision based on it.
4. Is Cloudflare Clairic available for on-premise or private cloud deployment? Currently, Clairic is only available as a hosted API on Cloudflare's Workers AI platform. However, Cloudflare has stated its intention to release the model weights openly, which could eventually allow for private deployment.
5. How accurate is Clairic's forecasting feature compared to a statistical model like ARIMA? Clairic's forecasting is a general-purpose tool for quick, integrated predictions from textual data. It is less accurate than a statistically rigorous, fine-tuned ARIMA or Prophet model fit to clean, granular time-series data but is far easier and faster to implement for ad-hoc or multi-series forecasting within a data processing workflow.