Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

fix(ingestion): honor auto-keyword/auto-question top_n sliders in extractor prompts (#18632)

euvre743491b8fb11
diff --git a/internal/ingestion/component/extractor.go b/internal/ingestion/component/extractor.go
index d2cdcf04cd..b5646e1fe9 100644
--- a/internal/ingestion/component/extractor.go
+++ b/internal/ingestion/component/extractor.go
@@ -126,6 +126,18 @@ func SetExtractorConcurrency(n int) {
}
}
+// extractorTopNPattern matches the {{ topn }} placeholder accepted in
+// keyword/question system prompts (same convention as rag/prompts/*.md).
+// It is replaced with the configured top_n so the count slider stays
+// authoritative even when a prompt was pre-filled by the frontend.
+var extractorTopNPattern = regexp.MustCompile(`\{\{\s*topn\s*\}\}`)
+
+// renderExtractorPrompt substitutes every {{ topn }} placeholder in the
+// system prompt with the configured extraction count.
+func renderExtractorPrompt(prompt string, topN int) string {
+ return extractorTopNPattern.ReplaceAllString(prompt, strconv.Itoa(topN))
+}
+
const (
autoKeywordPrompt = `## Role
You are a text analyzer.
@@ -134,7 +146,7 @@ You are a text analyzer.
Extract the most important keywords/phrases of a given piece of text content.
## Requirements
-- Summarize the text content, and give the top %d important keywords/phrases.
+- Summarize the text content, and give the top {{ topn }} important keywords/phrases.
- The keywords MUST be in the same language as the given piece of text content.
- The keywords are delimited by ENGLISH COMMA.
- Output keywords ONLY.`
@@ -146,7 +158,7 @@ You are a text analyzer.
Propose questions about a given piece of text content.
## Requirements
-- Understand and summarize the text content, and propose the top %d important questions.
+- Understand and summarize the text content, and propose the top {{ topn }} important questions.
- The questions SHOULD NOT have overlapping meanings.
- The questions SHOULD cover the main content of the text as much as possible.
- The questions MUST be in the same language as the given piece of text content.
@@ -629,10 +641,11 @@ func (c *ExtractorComponent) runAutoKeywords(ctx context.Context, db *gorm.DB, i
if topN <= 0 {
return nil
}
- systemPrompt := c.Param.Keywords.SystemPrompt
- if strings.TrimSpace(systemPrompt) == "" {
- systemPrompt = fmt.Sprintf(autoKeywordPrompt, topN)
+ systemPrompt := strings.TrimSpace(c.Param.Keywords.SystemPrompt)
+ if systemPrompt == "" {
+ systemPrompt = autoKeywordPrompt
}
+ systemPrompt = renderExtractorPrompt(systemPrompt, topN)
kwTemp := extractorTemperature
kwIn := extractorInputs{
llmID: in.llmID,
@@ -669,10 +682,11 @@ func (c *ExtractorComponent) runAutoQuestions(ctx context.Context, db *gorm.DB,
if topN <= 0 {
return nil
}
- systemPrompt := c.Param.Questions.SystemPrompt
- if strings.TrimSpace(systemPrompt) == "" {
- systemPrompt = fmt.Sprintf(autoQuestionPrompt, topN)
+ systemPrompt := strings.TrimSpace(c.Param.Questions.SystemPrompt)
+ if systemPrompt == "" {
+ systemPrompt = autoQuestionPrompt
}
+ systemPrompt = renderExtractorPrompt(systemPrompt, topN)
qTemp := extractorTemperature
qIn := extractorInputs{
llmID: in.llmID,
diff --git a/internal/ingestion/component/extractor_test.go b/internal/ingestion/component/extractor_test.go
index 7aca970979..5c7e579d2e 100644
--- a/internal/ingestion/component/extractor_test.go
+++ b/internal/ingestion/component/extractor_test.go
@@ -1712,6 +1712,103 @@ func TestExtractorCustomKeywordsAndQuestionsSystemPrompt(t *testing.T) {
}
}
+// TestExtractorTopNPlaceholderSubstitution verifies the {{ topn }} placeholder
+// in custom keyword/question system prompts is replaced with the configured
+// top_n, so the count slider stays authoritative when the frontend pre-fills
+// a prompt.
+func TestExtractorTopNPlaceholderSubstitution(t *testing.T) {
+ stub := withStubChatInvoker(t,
+ stubResponse{Content: "alpha, beta"},
+ stubResponse{Content: "q1?\nq2?"},
+ )
+
+ params := map[string]any{
+ "llm_id": "llm-1",
+ "keywords": map[string]any{
+ "top_n": 9,
+ "system_prompt": "Give the top {{ topn }} keywords.",
+ },
+ "questions": map[string]any{
+ "top_n": 7,
+ "system_prompt": "Propose {{topn}} questions.",
+ },
+ }
+
+ comp, err := NewExtractorComponent(params)
+ if err != nil {
+ t.Fatalf("NewExtractorComponent: %v", err)
+ }
+
+ in := map[string]any{
+ "chunks": []map[string]any{
+ {"content_with_weight": "Content text."},
+ },
+ }
+
+ if _, err := comp.Invoke(t.Context(), nil, in); err != nil {
+ t.Fatalf("Invoke: %v", err)
+ }
+
+ reqs := stub.requests
+ if len(reqs) != 2 {
+ t.Fatalf("expected 2 LLM calls, got %d", len(reqs))
+ }
+
+ if got := reqs[0].Messages[0].Content; got != "Give the top 9 keywords." {
+ t.Errorf("expected keywords prompt with top_n=9 substituted, got: %q", got)
+ }
+ if got := reqs[1].Messages[0].Content; got != "Propose 7 questions." {
+ t.Errorf("expected questions prompt with top_n=7 substituted, got: %q", got)
+ }
+}
+
+// TestExtractorDefaultPromptsRenderTopN verifies the built-in keyword/question
+// prompts interpolate the configured top_n when no custom prompt is set.
+func TestExtractorDefaultPromptsRenderTopN(t *testing.T) {
+ stub := withStubChatInvoker(t,
+ stubResponse{Content: "k1, k2"},
+ stubResponse{Content: "q1?"},
+ )
+
+ params := map[string]any{
+ "llm_id": "llm-1",
+ "keywords": map[string]any{"top_n": 4},
+ "questions": map[string]any{"top_n": 6},
+ }
+
+ comp, err := NewExtractorComponent(params)
+ if err != nil {
+ t.Fatalf("NewExtractorComponent: %v", err)
+ }
+
+ in := map[string]any{
+ "chunks": []map[string]any{
+ {"content_with_weight": "Content text."},
+ },
+ }
+
+ if _, err := comp.Invoke(t.Context(), nil, in); err != nil {
+ t.Fatalf("Invoke: %v", err)
+ }
+
+ reqs := stub.requests
+ if len(reqs) != 2 {
+ t.Fatalf("expected 2 LLM calls, got %d", len(reqs))
+ }
+
+ kwPrompt := reqs[0].Messages[0].Content
+ if !strings.Contains(kwPrompt, "top 4 important keywords/phrases") {
+ t.Errorf("expected built-in keywords prompt with top_n=4, got: %q", kwPrompt)
+ }
+ qPrompt := reqs[1].Messages[0].Content
+ if !strings.Contains(qPrompt, "top 6 important questions") {
+ t.Errorf("expected built-in questions prompt with top_n=6, got: %q", qPrompt)
+ }
+ if strings.Contains(kwPrompt, "{{") || strings.Contains(qPrompt, "{{") {
+ t.Errorf("expected no leftover placeholders, got keywords=%q questions=%q", kwPrompt, qPrompt)
+ }
+}
+
func TestExtractorDisabledSummarySkipsCall(t *testing.T) {
stub := withStubChatInvoker(t, stubResponse{Content: "Not expected"})
diff --git a/web/src/locales/bg.ts b/web/src/locales/bg.ts
index 5a49118918..978936f5ab 100644
--- a/web/src/locales/bg.ts
+++ b/web/src/locales/bg.ts
@@ -2186,7 +2186,7 @@ Task
Extract the most important keywords/phrases of a given piece of text content.
Requirements
-- Summarize the text content, and give the top 5 important keywords/phrases.
+- Summarize the text content, and give the top {{ topn }} important keywords/phrases.
- The keywords MUST be in the same language as the given piece of text content.
- The keywords are delimited by ENGLISH COMMA.
- Output keywords ONLY.`,
@@ -2194,10 +2194,10 @@ Requirements
You are a text analyzer.
Task
-Propose 3 questions about a given piece of text content.
+Propose {{ topn }} questions about a given piece of text content.
Requirements
-- Understand and summarize the text content, and propose the top 3 important questions.
+- Understand and summarize the text content, and propose the top {{ topn }} important questions.
- The questions SHOULD NOT have overlapping meanings.
- The questions SHOULD cover the main content of the text as much as possible.
- The questions MUST be in the same language as the given piece of text content.
diff --git a/web/src/locales/en.ts b/web/src/locales/en.ts
index a2bc4d9825..39d1373897 100644
--- a/web/src/locales/en.ts
+++ b/web/src/locales/en.ts
@@ -3361,7 +3361,7 @@ Task
Extract the most important keywords/phrases of a given piece of text content.
Requirements
-- Summarize the text content, and give the top 5 important keywords/phrases.
+- Summarize the text content, and give the top {{ topn }} important keywords/phrases.
- The keywords MUST be in the same language as the given piece of text content.
- The keywords are delimited by ENGLISH COMMA.
- Output keywords ONLY.`,
@@ -3369,10 +3369,10 @@ Requirements
You are a text analyzer.
Task
-Propose 3 questions about a given piece of text content.
+Propose {{ topn }} questions about a given piece of text content.
Requirements
-- Understand and summarize the text content, and propose the top 3 important questions.
+- Understand and summarize the text content, and propose the top {{ topn }} important questions.
- The questions SHOULD NOT have overlapping meanings.
- The questions SHOULD cover the main content of the text as much as possible.
- The questions MUST be in the same language as the given piece of text content.
diff --git a/web/src/locales/ko.ts b/web/src/locales/ko.ts
index 5af43427e7..612a598385 100644
--- a/web/src/locales/ko.ts
+++ b/web/src/locales/ko.ts
@@ -2679,7 +2679,7 @@ Task
Extract the most important keywords/phrases of a given piece of text content.
Requirements
-- Summarize the text content, and give the top 5 important keywords/phrases.
+- Summarize the text content, and give the top {{ topn }} important keywords/phrases.
- The keywords MUST be in the same language as the given piece of text content.
- The keywords are delimited by ENGLISH COMMA.
- Output keywords ONLY.`,
@@ -2687,10 +2687,10 @@ Requirements
You are a text analyzer.
Task
-Propose 3 questions about a given piece of text content.
+Propose {{ topn }} questions about a given piece of text content.
Requirements
-- Understand and summarize the text content, and propose the top 3 important questions.
+- Understand and summarize the text content, and propose the top {{ topn }} important questions.
- The questions SHOULD NOT have overlapping meanings.
- The questions SHOULD cover the main content of the text as much as possible.
- The questions MUST be in the same language as the given piece of text content.
diff --git a/web/src/locales/zh.ts b/web/src/locales/zh.ts
index b18fbe58c5..b7d5f0e4bb 100644
--- a/web/src/locales/zh.ts
+++ b/web/src/locales/zh.ts
@@ -2941,7 +2941,7 @@ Tokenizer 会根据所选方式将内容存储为对应的数据结构。`,
从给定的文本内容中提取最重要的关键词/短语。
要求
-- 总结文本内容,并给出最重要的5个关键词/短语。
+- 总结文本内容,并给出最重要的{{ topn }}个关键词/短语。
- 关键词必须与给定的文本内容使用相同的语言。
- 关键词之间用英文逗号分隔。
- 仅输出关键词。`,
@@ -2949,10 +2949,10 @@ Tokenizer 会根据所选方式将内容存储为对应的数据结构。`,
你是一名文本分析员。
任务
-针对给定的文本内容提出3个问题。
+针对给定的文本内容提出{{ topn }}个问题。
要求
-- 理解并总结文本内容,并提出最重要的3个问题。
+- 理解并总结文本内容,并提出最重要的{{ topn }}个问题。
- 问题的含义不应重叠。
- 问题应尽可能涵盖文本的主要内容。
- 问题必须与给定的文本内容使用相同的语言。
diff --git a/web/src/pages/agent/form/extractor-form/use-switch-prompt.ts b/web/src/pages/agent/form/extractor-form/use-switch-prompt.ts
index 1efe7b409c..00a90706a4 100644
--- a/web/src/pages/agent/form/extractor-form/use-switch-prompt.ts
+++ b/web/src/pages/agent/form/extractor-form/use-switch-prompt.ts
@@ -4,6 +4,11 @@ import { useTranslation } from 'react-i18next';
type SwitchPromptField = 'field_name' | 'sys_prompt' | 'prompts';
+// The Python canvas Extractor has no top_n input and sends sys_prompt to the
+// LLM verbatim, so its keyword and question prompts pin the {{ topn }}
+// placeholder to fixed counts.
+const pythonCanvasTopN: Record<string, number> = { keywords: 5, questions: 3 };
+
type SwitchPromptForm = {
getValues(name: 'field_name'): string;
setValue(
@@ -20,10 +25,17 @@ export function useSwitchPrompt(form: SwitchPromptForm) {
const setPromptValue = useCallback(
(field: SwitchPromptField, key: string, value: string) => {
- form.setValue(field, t(`flow.prompts.${key}.${value}`), {
- shouldDirty: true,
- shouldValidate: true,
- });
+ const prompt = t(`flow.prompts.${key}.${value}`);
+ form.setValue(
+ field,
+ Object.hasOwn(pythonCanvasTopN, value)
+ ? prompt.replace(/\{\{\s*topn\s*\}\}/g, String(pythonCanvasTopN[value]))
+ : prompt,
+ {
+ shouldDirty: true,
+ shouldValidate: true,
+ },
+ );
},
[form, t],
);