Anthropic Grounds Claude Outputs with New, Seamless Citations Feature

Anthropic is unveiling a new API feature for Claude—the company's next-generation AI assistant—which helps ground and contextualize Claude answers with source documentation. Now available for the Anthropic API, Google Cloud’s Vertex AI, Claude 3.5 Sonnet, and Claude 3.5 Haiku, Citations boost response trust by delivering detailed references to the exact source material used in its outputs.

Claude’s new Citations capability builds on the company’s commitment to developing inherently trustworthy, steerable models. Before the launch of Citations, developers had to create complex prompts to instruct Claude to cite its sources. Now, developers can simply add source documents to Claude’s context window; during a query, Claude automatically cites its output with the sources that it derived its information from.

According to Tarun Amasa, CEO, Endex, a user of Claude, "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.”

Amasa’s testament is further affirmed by Anthropic’s internal evaluations of Citations, which found that it outperformed most custom implementations and increased recall accuracy by up to 15%.

Under the hood, Citations works by chunking the user-provided source documents (PDF documents and plain text files) into sentences, which are then passed through the model with the user’s query and additional context provided by the user. Users are also allowed to chunk their own documents.

Citations does not require file storage and seamlessly integrates with the Messages API, enabling users to easily deliver contextual, source-transparent responses through Claude.

Citations can be applied to a multitude of use cases, including document summaries paired with links to original sources; complex Q&As that span a large corpus of documentation which can be traced back to relevant texts; and customer support systems that reference multiple product manuals, support tickets, and more.

Another customer testimony from Jake Heller, head of product, CoCounsel, at Thomson Reuters, echoed Amasa’s testament.

“For CoCounsel to be trustworthy and immediately useful for practicing attorneys, it needs to cite its work. We first built this ourselves, but it was really hard to build and maintain. That's why we were excited to test out Anthropic’s Citations functionality,” said Heller. “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. The Citations feature will enable us to build an even more accurate and thorough AI assistant for lawyers.”

Citations utilizes Anthropic’s standard token-based pricing model; it may use additional input tokens to process documents but does not require payment for output tokens that return the quoted text itself, according to the company. 

To learn more about Citations and Claude, please visit https://www.anthropic.com/.