On January 23, 2025, Anthropic announced the launch of a new feature named "Citations" for its Claude API. This enhancement enables Claude to anchor its responses in specific source documents, significantly enhancing the verifiability and trustworthiness of the AI-generated outputs. The Citations feature is now generally accessible on both the Anthropic API and Google Cloud’s Vertex AI platform.
All Claude models are inherently designed to prioritize trustworthiness and steerability. The introduction of Citations addresses a vital requirement in artificial intelligence applications: the ability to verify the sources of AI-generated content. Previously, developers faced challenges using complex prompts to instruct Claude to include source information, which often led to inconsistent results and increased times for prompt engineering and testing. The Citations feature simplifies this process by allowing users to add source documents directly into the context window. When queried, Claude will automatically provide citations for any claims derived from these documents.
Internal assessments highlight that Claude's citation capabilities surpass most custom solutions, enhancing recall accuracy by up to 15%.
The Citations feature opens new avenues for developers looking to implement AI solutions with heightened accountability, including:
When activated, the Citations feature processes user-provided source materials, such as PDF and plain text files, by segmenting them into sentences. These segments, alongside user-provided context, are relayed to the model along with the user inquiry. Users also have the option to supply pre-defined segments.
Claude utilizes this data to deliver responses embedded with precise citations, referencing original documents to reduce misinformation and enhance reliability. This method integrates seamlessly with the Messages API by bypassing the need for file storage and offering greater flexibility.
The Citations feature adheres to Anthropic’s standard token-based pricing model. While it may require more input tokens for document processing, users are not charged for output tokens related to the cited text itself.
Thomson Reuters employs Claude for its AI platform, CoCounsel, assisting legal and tax professionals in synthesizing expertise and providing comprehensive advice. Jake Heller, Head of Product at CoCounsel, stated, "For CoCounsel to be trustworthy and immediately useful for practicing attorneys, it needs to cite its work. We initially developed this internally, but maintaining it proved challenging. The integration of Anthropic’s Citations feature has simplified building, maintaining, and deploying reliable citation capabilities for our users, reducing hallucination risks and bolstering trust in AI-generated content. This functionality will enable us to develop an even more precise AI assistant for legal professionals."
Endex harnesses Claude to drive an Autonomous Agent for financial institutions. According to Tarun Amasa, CEO of Endex, "Anthropic's Citations reduced our source hallucinations and formatting issues significantly, improving reference quality by 20%. This improvement has negated the need for elaborate prompt engineering focused on references, enhancing our accuracy in performing intricate, multi-layered financial analysis."
Citations are now available for Claude 3.5 Sonnet and Claude 3.5 Haiku. Users can access detailed documentation to begin integrating this feature effectively into their systems.
For further information and support, visit our website or contact our team.
