{"id":708,"date":"2025-01-25T03:45:16","date_gmt":"2025-01-25T03:45:16","guid":{"rendered":"https:\/\/www.batteryone.co\/blog\/?p=708"},"modified":"2025-01-25T03:45:16","modified_gmt":"2025-01-25T03:45:16","slug":"anthropics-new-citations-feature-for-claude-a-step-toward-enhanced-accuracy-and-trust","status":"publish","type":"post","link":"https:\/\/www.batteryone.co\/blog\/archives\/708","title":{"rendered":"Anthropic&#8217;s New Citations Feature for Claude: A Step Toward Enhanced Accuracy and Trust"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\">On Thursday,\u00a0<strong>Anthropic<\/strong>\u00a0int<strong>Anthropic&#8217;s New Citations Feature for Claude: A Step Toward Enhanced Accuracy and Trust<\/strong>oduced a new API feature called\u00a0<strong>Citations<\/strong>, designed to help Claude models avoid the common issue of\u00a0<strong>confabulations<\/strong>\u00a0or\u00a0<strong>hallucinations<\/strong>\u00a0by linking their responses directly to source documents. This new tool allows developers to upload documents (PDFs and plaintext files) into Claude\u2019s context window, enabling the AI to automatically cite specific passages used in its answers.<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\" src=\"https:\/\/cdn.techinasia.com\/wp-content\/uploads\/2024\/11\/1732598623_shutterstock_2333831851-750x429.jpg\" alt=\"\"\/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">How Citations Work<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">When the&nbsp;<strong>Citations<\/strong>&nbsp;feature is enabled, Anthropic\u2019s API processes the user-provided documents by&nbsp;<strong>chunking<\/strong>&nbsp;them into&nbsp;<strong>sentences<\/strong>. These sentences, along with any additional context provided by the user, are then passed to Claude with the user\u2019s query. The model will use this context to generate answers, referencing the specific passages it utilized.<br>As&nbsp;<strong>Anthropic<\/strong>&nbsp;explains, this feature has a wide range of potential applications, including:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Summarizing\u00a0<strong>case files<\/strong>\u00a0with\u00a0<strong>source-linked key points<\/strong><\/li>\n\n\n\n<li>Answering questions from\u00a0<strong>financial documents<\/strong>\u00a0with\u00a0<strong>traced references<\/strong><\/li>\n\n\n\n<li>Powering\u00a0<strong>support systems<\/strong>\u00a0that cite specific\u00a0<strong>product documentation<\/strong><\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">In its internal testing, Anthropic found that&nbsp;<strong>Citations<\/strong>&nbsp;improved&nbsp;<strong>recall accuracy<\/strong>&nbsp;by up to&nbsp;<strong>15 percent<\/strong>&nbsp;compared to custom citation systems built by users within prompts. Although a 15% improvement may seem modest,&nbsp;<strong>Simon Willison<\/strong>, a well-known AI researcher, highlighted the feature\u2019s significance due to its integration of&nbsp;<strong>Retrieval-Augmented Generation (RAG)<\/strong>&nbsp;techniques. RAG is an approach where, after retrieving relevant portions of documents, the model generates an answer that includes these fragments, ensuring more accurate and contextually relevant responses.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">&gt;&gt;&gt;<a href=\"https:\/\/www.batteryone.co\/detail\/1747092\/Li318991PVYTL\">Li318991PVYTL<\/a>&nbsp;Replacement Battery for&nbsp;<a href=\"https:\/\/www.batteryone.co\/brand\/3\/Blackview\">Blackview<\/a>&nbsp;Mega 1<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Willison pointed out that while using citations helps&nbsp;<strong>verify accuracy<\/strong>, building a system that does so consistently is challenging. However, Citations seems to be a step in the right direction by integrating RAG techniques directly into the model. As Willison writes on his blog,&nbsp;<strong>\u201cThe core of the Retrieval Augmented Generation (RAG) pattern is to take a user\u2019s question, retrieve portions of documents that might be relevant to that question and then answer the question by including those text fragments in the context provided to the LLM.\u201d<\/strong>&nbsp;This setup helps mitigate the risk of models answering based on outdated or incorrect training data.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Early adopters<\/strong>&nbsp;of Citations have reported&nbsp;<strong>promising results<\/strong>. For instance,&nbsp;<strong>Thomson Reuters<\/strong>, which uses Claude to power its&nbsp;<strong>CoCounsel<\/strong>&nbsp;legal AI reference platform, expressed excitement about the potential of Citations to&nbsp;<strong>minimize hallucinations<\/strong>&nbsp;and&nbsp;<strong>increase trust<\/strong>&nbsp;in AI-generated content. In addition,&nbsp;<strong>Endex<\/strong>, a&nbsp;<strong>financial technology company<\/strong>, shared that&nbsp;<strong>Citations<\/strong>&nbsp;helped&nbsp;<strong>reduce source confabulations from 10% to zero<\/strong>, while also&nbsp;<strong>increasing the number of references per response<\/strong>&nbsp;by&nbsp;<strong>20%<\/strong>.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Despite these promising results, Anthropic and other developers caution that relying on any language model to&nbsp;<strong>accurately relay reference information<\/strong>&nbsp;still carries inherent risks, especially when the technology is still evolving. The company itself emphasizes that this capability should be viewed as part of a broader effort to improve the&nbsp;<strong>reliability<\/strong>&nbsp;and&nbsp;<strong>trustworthiness<\/strong>&nbsp;of AI-generated content, and it may need further development and testing in real-world applications.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">&gt;&gt;&gt;<a href=\"https:\/\/www.batteryone.co\/detail\/1747091\/01AC366\">01AC366<\/a>&nbsp;Replacement Battery for&nbsp;<a href=\"https:\/\/www.batteryone.co\/brand\/13\/IBM\">IBM<\/a>&nbsp;Storwize V5000 V5010 V5030 Gen2<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Pricing and Availability<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Citations is available for&nbsp;<strong>Claude 3.5 Sonnet<\/strong>&nbsp;and&nbsp;<strong>Claude 3.5 Haiku models<\/strong>&nbsp;through both the&nbsp;<strong>Anthropic API<\/strong>&nbsp;and&nbsp;<strong>Google Cloud&#8217;s Vertex AI platform<\/strong>. According to&nbsp;<strong>Anthropic\u2019s token-based pricing<\/strong>, quoting text from a source document won\u2019t count toward the output token costs. For instance, sourcing a&nbsp;<strong>100-page document<\/strong>&nbsp;as a reference would cost around&nbsp;<strong>$0.30<\/strong>&nbsp;with&nbsp;<strong>Claude 3.5 Sonnet<\/strong>&nbsp;or&nbsp;<strong>$0.08<\/strong>&nbsp;with&nbsp;<strong>Claude 3.5 Haiku<\/strong>..<\/p>\n","protected":false},"excerpt":{"rendered":"<p>On Thursday,\u00a0Anthropic\u00a0intAnthropic&#8217;s New Citations Feature for Claude: A Step Toward Enhanced Accuracy and Trustoduced a new API feature called\u00a0Citations, designed to help Claude models avoid the common issue of\u00a0confabulations\u00a0or\u00a0hallucinations\u00a0by linking their responses directly to source documents. This new tool allows developers to upload documents (PDFs and plaintext files) into Claude\u2019s context window, enabling the AI [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[68],"class_list":["post-708","post","type-post","status-publish","format-standard","hentry","category-news","tag-ai"],"_links":{"self":[{"href":"https:\/\/www.batteryone.co\/blog\/wp-json\/wp\/v2\/posts\/708","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.batteryone.co\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.batteryone.co\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.batteryone.co\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.batteryone.co\/blog\/wp-json\/wp\/v2\/comments?post=708"}],"version-history":[{"count":1,"href":"https:\/\/www.batteryone.co\/blog\/wp-json\/wp\/v2\/posts\/708\/revisions"}],"predecessor-version":[{"id":709,"href":"https:\/\/www.batteryone.co\/blog\/wp-json\/wp\/v2\/posts\/708\/revisions\/709"}],"wp:attachment":[{"href":"https:\/\/www.batteryone.co\/blog\/wp-json\/wp\/v2\/media?parent=708"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.batteryone.co\/blog\/wp-json\/wp\/v2\/categories?post=708"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.batteryone.co\/blog\/wp-json\/wp\/v2\/tags?post=708"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}