Willison notes that although quoting sources helps to verify accuracy, building a system that does this well “can be quite difficult”, but Citations seems to be a step in the right direction by RAG options directly in directly in to build in the model.
Apparently that possibility is not new. Alex Albert van Anthropic wrote about X: “Under the hood, Claude is trained in quoting sources. With Citations we expose this opportunity to developers. To use citations, users can have a new parameter” Citations: {enabled: True} ” Pass on to each document type that they send via the API. “
Early Adopter reports promising results
The company has released Citations for the Claude 3.5 Sonnet and Claude 3.5 Haiku models via both the Anthropic API and the Vertex AI platform from Google Cloud, but it is apparently already used in practice.
Anthropic says that Thomson Reuters, that Claude uses to drive his legal AI reference platform Cocounsel, look forward to using Citations in a way that helps to minimize the risk of hallucinations, but also reinforces the trust in content. “
In addition, Financial Technology Company Endex told Anthropic that Citations has reduced their source confabulations from 10 percent to zero, while the number of references per reaction has increased by 20 percent, according to CEO Tarun Amasa.
Despite these claims, it is still a risk to rely on an LLM to accurately pass on reference information until the technology has been studied in a more stupid and proven in the field.
Anthropic will charge users the standard on tokens -based prices, although the quoted text in responses does not count for the outputtoken costs. Purchasing a 100-page document would cost around $ 0.30 with Claude 3.5 Sonnet or $ 0.08 with Claude 3.5 Haiku, according to the standard API prices of Anthropic.
