Anthropic has introduced "Citations," a new API feature designed to improve the reliability and accuracy of its Claude AI models.
This functionality allows Claude to ground its answers in specific source documents, providing explicit references to the exact lines and chapters used in its responses. Citations reduce hallucinations linking generated content to its roots, hence increasing the dependability of AI-derived information.
Citations, which are available for Claude 3.5 Sonnet and Claude 3.5 Haiku via Anthropic's API and Google's Cloud's Vertex AI, solve the difficulty in AI development: verifying the sources of AI outputs. Developers can upload documents, which are then analyzed into pieces, allowing Claude to quote assertions derived from these files.
This functionality greatly minimizes the requirement for substantial prompt engineering, saving developers time and effort while enhancing accuracy. Internal evaluation indicates Citations increases recall accuracy by up to 15%.
Citations are especially valuable for use cases like document summarization, sophisticated Q&A, and customer assistance, which require source-backed replies.
For example, Thomson Reuters uses Citations in its CoCounsel AI platform for legal and tax professionals to ensure accurate and trustworthy references. Similarly, Endex, a financial services platform, found a 20% increase in reference accuracy and the elimination of hallucinations in financial research outputs.
While Citations charges additional fees for token consumption, its pricing is competitive, with a 100-page document costing as little as $0.08 when using Claude 3.5 Haiku. This affordability, combined with the feature's dependability, makes compiling a choice for developers looking to boost transparency and trust in AI systems.
Anthropic's Citations marks a significant step towards making AI output more verifiable, positioning the company as a leader in accountable AI development.