Anthropic Launches Citations for Accurate AI Responses

Anthropic has introduced Citations, a new API feature designed to improve the credibility of responses generated by its Claude models. With this enhancement, Claude can now reference specific sentences and passages from source documents, offering users clear and verifiable links to the information that informs its answers.

Citations is now widely available on the Anthropic API and Google Cloud’s Vertex AI.

Strengthening AI Trust Through Citations

Claude models are built with a strong focus on reliability and adaptability. The addition of Citations addresses a key challenge in AI development: providing clear evidence to back AI-generated answers.

Previously, developers needed intricate prompts to coax source details from Claude, which often resulted in variable outcomes and extensive time spent on prompt engineering. With Citations, users can include source materials directly in the context window. Claude will then automatically cite information drawn from these documents in its responses.

Improved Accuracy:

According to Anthropic’s internal testing, the Citations feature outperforms most customized solutions, increasing recall accuracy by up to 15%.

Practical Applications of Citations

Citations unlock new possibilities for developers, particularly in use cases requiring high levels of accountability and accuracy, such as:

• Document Summarization: Generate concise overviews of lengthy texts, like legal or business files, with each key point traced back to its original source.

• Complex Q&A: Deliver in-depth answers to queries across extensive document repositories, such as annual or financial reports, with every element supported by specific references.

• Customer Support: Create support systems that draw from multiple resources, including product guides, support tickets, and FAQs, while always citing the exact source of the information provided.

How Citations Work

When enabled, the Citations feature processes user-supplied documents (e.g., PDFs or text files) by dividing them into manageable chunks, such as individual sentences. These chunks, along with user-provided context, are sent to Claude along with the query. Users can also pre-chunk their documents if needed.

Claude generates responses by analyzing the query and the provided materials, embedding citations for any claims that rely on those sources. This approach minimizes hallucinations and enhances the reliability of the generated content.

Citations seamlessly integrate with the Messages API and do not require external storage for documents, making it an efficient solution for developers.

Transparent and Flexible Pricing

Citations operates under Anthropic’s standard token-based pricing model. While processing source materials may consume additional input tokens, users are not charged for the output tokens that contain quoted text.

Customer Success Stories

Thomson Reuters: Thomson Reuters uses Claude to power CoCounsel, an AI platform designed for legal and tax professionals. Jake Heller, Head of Product for CoCounsel, highlighted the benefits of Citations:

“It makes citing and linking to primary sources much easier to build, maintain, and deploy to our users. This capability not only helps minimize hallucination risk but also strengthens trust in AI-generated content.”

Endex: Endex leverages Claude for financial firms through an Autonomous Agent. CEO Tarun Amasa shared how Citations enhanced their operations:

“With Anthropic's Citations, we reduced source hallucinations and formatting issues from 10% to 0% and saw a 20% increase in references per response. This removed the need for elaborate prompt engineering around references and improved our accuracy when conducting complex, multi-stage financial research.”

Getting Started

The Citations feature is available for Claude 3.5 Sonnet and Claude 3.5 Haiku. Developers can begin by exploring the Citations documentation.

What This Means

By enabling Claude to cite sources, Anthropic is raising the bar for AI transparency and trustworthiness. By directly linking claims to sources, Anthropic has provided developers with a powerful tool for building applications that require precision and accountability, particularly in legal, financial, and customer support domains.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.