On Thursday, Anthropic unveiled a new application programming interface (API) feature designed to enhance the accuracy of responses generated by its artificial intelligence (AI) models. The new functionality, named Citations, allows developers to tie the output of Claude AI models directly to specific source documents. This initiative aims to bolster the reliability of the information produced by these models. Companies such as Thomson Reuters, which utilizes the feature for its CoCounsel platform, and Endex have already integrated it, with no additional costs involved.
Anthropic Introduces a New Grounding Feature
Generative AI models can often produce errors and engage in hallucinations due to the extensive datasets they process to answer user inquiries. The inclusion of web searches complicates matters further, making it difficult for large language models (LLMs) to filter out incorrect information since they rely on basic retrieval-augmented generation (RAG) techniques.
To enhance the accuracy and reliability of their outputs, AI companies often limit the data access of their LLMs. Examples include features like Gemini in Google Docs, AI writing assistance tools on Samsung and Apple devices, and PDF analysis capabilities in Adobe Acrobat. However, this restricted approach is challenging to implement within an API framework given the diverse data needs of developers building a variety of tools.
To address these challenges, Anthropic has introduced the Citations feature, which allows the Claude models to ground their outputs in specific source documents. Detailed in a recent newsroom post, this feature enables Claude to reference the exact paragraphs and sentences from which it derives information, making the AI-generated content more verifiable and trustworthy.
Users can incorporate source documents directly into the context window, and Claude will automatically include citations in its responses where applicable. This development aims to eliminate the need for developers to rely on complex prompts to request source information—a method that Anthropic recognizes as inefficient and inconsistent.
According to Anthropic, the Citations feature will facilitate the creation of AI solutions focused on document summarization, complex query resolutions based on lengthy texts, and customer support systems.
Importantly, the Citations feature operates under Anthropic’s standard token-based pricing model. Users will not incur charges for output tokens that deliver quoted text, though there may be additional costs for input tokens used in processing the source documents. Currently, Citations is available for the Claude 3.5 Sonnet and Claude 3.5 Haiku models.
