Anthropic Introduces Citations Feature to Enhance AI Model Accuracy and Reliability

Anthropic Unveils Citations Feature for Enhanced AI Accuracy

Anthropic, a leading AI company, has introduced a new API feature called Citations, aimed at improving the accuracy and reliability of its Claude family of AI models. Announced on Thursday, this feature allows developers to ground AI-generated responses in source documents, potentially reducing the occurrence of AI hallucinations and confabulations 1

1

.

How Citations Works

The Citations feature processes user-provided source documents by chunking them into sentences. These chunks, along with user-provided context, are then passed to the model along with the user's query. This approach enables Claude to automatically cite specific passages it uses to generate answers 1

1

.

Developers can add source files to have models automatically cite claims inferred from those files. The feature is particularly useful in document summarization, Q&A, and customer support applications 2

2

.

Availability and Pricing

Citations is currently available for Claude 3.5 Sonnet and Claude 3.5 Haiku models through both Anthropic's API and Google's Vertex AI platform 2

2

. The feature uses Anthropic's standard token-based pricing model, with no additional cost for output tokens returning quoted text. However, there might be extra charges for processing source documents 3

3

.

Potential Impact and Use Cases

Anthropic suggests several potential use cases for Citations, including:

1. Summarizing case files with source-linked key points
2. Answering questions across financial documents with traced references
3. Powering support systems that cite specific product documentation 1

   1

The company claims that in internal testing, the feature improved recall accuracy by up to 15 percent compared to custom citation implementations created by users within prompts 1

1

.

Industry Perspective

AI researchers, such as Simon Willison, have shown interest in the Citations feature due to its fundamental integration of Retrieval Augmented Generation (RAG) techniques. Willison explains that while RAG patterns generally work well, there is still a risk of models answering based on other information from their training data or hallucinating incorrect details 1

1

.

The introduction of Citations addresses a common challenge in generative AI models, which are prone to errors and hallucinations due to the massive datasets they process. By restricting data access and providing verifiable sources, Anthropic aims to make AI-generated responses more trustworthy and reliable 3

3

.