Anthropic launched a new application programming interface (API) feature on Thursday called Citations, designed to enhance the reliability and precision of responses produced by its artificial intelligence (AI) models. This feature enables developers to tie the output generated by the Claude family of AI models to source documents, thereby ensuring the responses are grounded in verifiable information. Notably, companies such as Thomson Reuters, which utilizes it in the CoCounsel platform, and Endex have already gained access to this capability at no additional cost.
Anthropic Introduces a New Grounding Feature
Generative AI models often face challenges related to inaccuracies and information hallucination, primarily due to their reliance on extensive datasets to generate user responses. Incorporating web searches complicates matters further, as large language models (LLMs) utilize basic retrieval-augmented generation (RAG) techniques to fetch information, making it difficult to ensure accuracy.
To tackle these accuracy concerns, AI firms frequently limit data access for LLMs, enabling them to provide more reliable outputs. Examples of such specialized tools include Gemini within Google Docs, AI-powered writing assistants on Samsung and Apple devices, and PDF analysis tools from Adobe Acrobat. However, the creation of such layers becomes challenging in an API context, where developers must accommodate a diverse array of data needs.
In response to this issue, Anthropic has rolled out the Citations feature for its API. According to a newsroom announcement, this feature allows Claude to reference source documents directly in its responses. This capability enables Claude AI models to identify and cite the exact paragraphs and sentences from which information is derived, thereby enhancing the verifiability and trustworthiness of the AI-generated content.
Users can now include source documents within the context window, prompting Claude to automatically cite the sources in its output as it leverages information from them. This development alleviates the need for developers to create complex prompts to request source information, an approach that the company recognized as inconsistent and cumbersome.
Anthropic asserts that the Citations feature will empower developers to create AI solutions for various applications, including document summarization, systems that can answer intricate queries from lengthy documents, and customer support platforms.
Remarkably, the company noted that Citations operates under its standard token-based pricing model, meaning users will not incur charges for output tokens related to quoted text. However, additional fees may apply for input tokens required to process the source documents. Currently, the Citations feature is available for the Claude 3.5 Sonnet and Claude 3.5 Haiku models.
