Claude 能夠在回答有關文件的問題時提供詳細引用,幫助您追蹤和驗證回應中的資訊來源。
所有活躍模型都支援引用,除了 Haiku 3。
Claude Sonnet 3.7 的引用
與其他 Claude 模型相比,Claude Sonnet 3.7 在沒有用戶更明確指示的情況下,可能較不傾向於進行引用。在使用 Claude Sonnet 3.7 進行引用時,我們建議在 user 回合中包含額外的指示,例如 "使用引用來支持您的答案。"
我們也觀察到,當模型被要求結構化其回應時,除非明確告知在該格式內使用引用,否則不太可能使用引用。例如,如果模型被要求在其回應中使用 <result> 標籤,您應該添加類似 "始終在您的答案中使用引用,即使在 <result> 標籤內也是如此。" 的內容
請使用此表單分享您對引用功能的回饋和建議。
以下是如何在 Messages API 中使用引用的範例:
curl https://api.anthropic.com/v1/messages \
-H "content-type: application/json" \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-d '{
"model": "claude-sonnet-4-5",
"max_tokens": 1024,
"messages": [
{
"role": "user",
"content": [
{
"type": "document",
"source": {
"type": "text",
"media_type": "text/plain",
"data": "The grass is green. The sky is blue."
},
"title": "My Document",
"context": "This is a trustworthy document.",
"citations": {"enabled": true}
},
{
"type": "text",
"text": "What color is the grass and sky?"
}
]
}
]
}'與基於提示的方法比較
與基於提示的引用解決方案相比,引用功能具有以下優勢:
cited_text 不計入您的輸出代幣。cited_text,引用保證包含對提供文件的有效指標。透過以下步驟將引用與 Claude 整合:
提供文件並啟用引用
文件處理
Claude 提供引用回應
自動分塊 vs 自訂內容
預設情況下,純文字和 PDF 文件會自動分塊為句子。如果您需要更多對引用粒度的控制(例如,對於項目符號或轉錄),請改用自訂內容文件。有關更多詳細資訊,請參閱文件類型。
例如,如果您希望 Claude 能夠從您的 RAG 塊中引用特定句子,您應該將每個 RAG 塊放入純文字文件中。否則,如果您不希望進行任何進一步的分塊,或者如果您想要自訂任何額外的分塊,您可以將 RAG 塊放入自訂內容文件中。
source 內容中找到的文字可以被引用。title 和 context 是可選欄位,將傳遞給模型但不用於引用內容。title 長度有限,因此您可能會發現 context 欄位在儲存任何文件元資料作為文字或字串化 json 時很有用。content 列表進行 0 索引,具有排他性結束索引。cited_text 欄位是為了方便而提供的,不計入輸出代幣。cited_text 也不計入輸入代幣。引用與其他 API 功能配合使用,包括提示快取、代幣計數和批次處理。
引用和提示快取可以有效地一起使用。
在回應中生成的引用塊不能直接快取,但它們引用的來源文件可以被快取。為了優化效能,將 cache_control 應用於您的頂層文件內容塊。
在此範例中:
cache_control 進行快取我們支援三種引用文件類型。文件可以直接在訊息中提供(base64、文字或 URL)或透過檔案 API上傳並通過 file_id 引用:
| 類型 | 最適合 | 分塊 | 引用格式 |
|---|---|---|---|
| 純文字 | 簡單文字文件、散文 | 句子 | 字符索引(0 索引) |
| 包含文字內容的 PDF 檔案 | 句子 | 頁碼(1 索引) | |
| 自訂內容 | 列表、轉錄、特殊格式、更細粒度的引用 | 無額外分塊 | 塊索引(0 索引) |
不支援 .csv、.xlsx、.docx、.md 和 .txt 檔案作為文件塊。將這些轉換為純文字並直接包含在訊息內容中。請參閱使用其他檔案格式。
純文字文件會自動分塊為句子。您可以內聯提供它們或通過其 file_id 引用:
PDF 文件可以作為 base64 編碼資料或通過 file_id 提供。PDF 文字被提取並分塊為句子。由於尚不支援圖像引用,掃描文件且不包含可提取文字的 PDF 將無法引用。
自訂內容文件讓您控制引用粒度。不進行額外分塊,塊根據提供的內容塊提供給模型。
{
"type": "document",
"source": {
"type": "content",
"content": [
{"type": "text", "text": "First chunk"},
{"type": "text", "text": "Second chunk"}
]
},
"title": "Document Title", # 可選
"context": "Context about the document that will not be cited from", # 可選
"citations": {"enabled": True}
}當啟用引用時,回應包含多個帶有引用的文字塊:
{
"content": [
{
"type": "text",
"text": "According to the document, "
},
{
"type": "text",
"text": "the grass is green",
"citations": [{
"type": "char_location",
"cited_text": "The grass is green.",
"document_index": 0,
"document_title": "Example Document",
"start_char_index": 0,
"end_char_index": 20
}]
},
{
"type": "text",
"text": " and "
},
{
"type": "text",
"text": "the sky is blue",
"citations": [{
"type": "char_location",
"cited_text": "The sky is blue.",
"document_index": 0,
"document_title": "Example Document",
"start_char_index": 20,
"end_char_index": 36
}]
},
{
"type": "text",
"text": ". Information from page 5 states that ",
},
{
"type": "text",
"text": "water is essential",
"citations": [{
"type": "page_location",
"cited_text": "Water is essential for life.",
"document_index": 1,
"document_title": "PDF Document",
"start_page_number": 5,
"end_page_number": 6
}]
},
{
"type": "text",
"text": ". The custom document mentions ",
},
{
"type": "text",
"text": "important findings",
"citations": [{
"type": "content_block_location",
"cited_text": "These are important findings.",
"document_index": 2,
"document_title": "Custom Content Document",
"start_block_index": 0,
"end_block_index": 1
}]
}
]
}對於串流回應,我們添加了一個 citations_delta 類型,其中包含要添加到當前 text 內容塊上的 citations 列表的單個引用。
import anthropic
client = anthropic.Anthropic()
# 長文件內容(例如,技術文件)
long_document = "This is a very long document with thousands of words..." + " ... " * 1000 # 最小可快取長度
response = client.messages.create(
model="claude-sonnet-4-5",
max_tokens=1024,
messages=[
{
"role": "user",
"content": [
{
"type": "document",
"source": {
"type": "text",
"media_type": "text/plain",
"data": long_document
},
"citations": {"enabled": True},
"cache_control": {"type": "ephemeral"} # 快取文件內容
},
{
"type": "text",
"text": "What does this document say about API features?"
}
]
}
]
)