Anthropic adds source citations to Claude models

Anthropic has rolled out "Citations," a new API feature that helps Claude automatically add source references to its responses, making AI-generated content more reliable and trustworthy.

The company just made Citations available through its API for Claude 3.5's Sonnet and Haiku models. According to Anthropic, this built-in citation system works better than most custom solutions, improving citation accuracy by about 15 percent.

The system breaks down source documents like PDFs and text files into individual sentences. It then combines these text fragments with the user's context and question before processing everything.

Before Citations, developers had to write complicated prompts just to get Claude to include basic source information. According to Anthropic, this manual approach often produced inconsistent results and forced developers to spend a lot of time fine-tuning their prompts.

The new feature uses Anthropic's standard token-based pricing system, and users don't have to pay extra for the quoted text in Claude's responses.

Mitigating source hallucinations through citations

Thomson Reuters plans to use Citations to make its AI legal platform CoCounsel more trustworthy, while financial firm Endex says it has already seen significant improvements with its AI financial agent.

The company reports that "source hallucinations" and formatting issues have dropped from 10 percent to zero, while the number of references in each response has increased by 20 percent. Both companies say that Citations makes their prompt engineering work easier.

Anthropic says Citations is great for document summaries, from summarizing documents to answering complex questions across large collections of files. The feature could also improve customer service by helping AI systems better understand product manuals and FAQs, the company says.