Anthropic Launches “Citations” API for More Trustworthy responses from Claude

Anthropic Launches “Citations” API for More Trustworthy responses from Claude

Anthropic has rolled out a new feature called "Citations" which pretty much does what the name suggests—provide refernces to any content that it uses from source documents. It allows developers to ground AI answers directly in user-provided documents, such as emails or PDFs. The feature is live on Anthropic’s API and Google Cloud’s Vertex AI.

Key Points:

  • Citations links AI-generated outputs to specific sentences and passages in source documents.
  • The feature simplifies development workflows by eliminating the need for custom prompt engineering.
  • Internal evaluations show Citations improves recall accuracy by up to 15% compared to custom implementations.
  • Supports Claude 3.5 Sonnet and Haiku on both Anthropic’s API and Vertex AI.

Citations aims to address a critical challenge in AI: the ability to verify the origins of AI-generated information. This builds on Claude’s design philosophy of trustworthiness and steerability, making it particularly suited for applications where accuracy and transparency are paramount.

Traditionally, developers relied on intricate prompts to embed citations within AI outputs. This approach often led to inconsistent results and a high demand for prompt engineering. Citations streamlines this process by allowing developers to attach source documents directly in the API’s context window. Claude automatically references the relevant sentences and sections from these sources in its responses, reducing the risk of hallucinations and improving accountability.

For example, Citations can:

  • Summarize long documents, linking each summary point back to its source.
  • Provide detailed answers to queries about complex documents like legal case files or financial statements.
  • Create customer support systems that pull answers from product manuals and FAQs while citing exact sources.

Citations processes documents by breaking them into “chunks,” such as sentences, which are then passed to the AI alongside the user query. This ensures responses cite specific content, formatted to reference document pages, text ranges, or custom content blocks. While this adds slight token costs for input, cited text is not counted in output tokens, offering developers cost efficiencies.

Early adopters like Thomson Reuters and Endex highlight the potential of Citations in professional settings:

  • Thomson Reuters integrated Citations into its CoCounsel AI platform, boosting trust and reducing hallucination risks in legal and tax advisory services.
  • Endex, a financial AI provider, reported a 20% increase in reference accuracy and eliminated citation formatting errors, streamlining multi-stage financial research.

Citations is expected to bolster Claude’s appeal among developers seeking robust, accountable AI integrations. Its design also complements other API features like token counting and batch processing, offering a seamless experience for a range of use cases.

Chris McKay is the founder and chief editor of Maginative. His thought leadership in AI literacy and strategic AI adoption has been recognized by top academic institutions, media, and global brands.

Let’s stay in touch. Get the latest AI news from Maginative in your inbox.

Subscribe