feat(summary): enhance document summarization capabilities and UI
- Added max_input_chars configuration to limit input size for summary generation. - Updated generate_summary.yaml to clarify summary generation steps and requirements. - Introduced a new summary section in the document content view to display generated summaries or loading states. - Refactored doc-content.vue to utilize computed properties for improved performance and readability. - Enhanced knowledge base hooks to include description and summary status fields for better data handling. - Updated internationalization files to include new summary-related labels in multiple languages.
This commit is contained in:
@@ -1910,7 +1910,10 @@ func (s *knowledgeService) processChunks(ctx context.Context,
|
||||
logger.GetLogger(ctx).Infof("processChunks successfully")
|
||||
}
|
||||
|
||||
// GetSummary generates a summary for knowledge content using an AI model
|
||||
// defaultMaxInputChars is the default maximum characters used as input for summary generation.
|
||||
const defaultMaxInputChars = 16384
|
||||
|
||||
// getSummary generates a summary for knowledge content using an AI model
|
||||
func (s *knowledgeService) getSummary(ctx context.Context,
|
||||
summaryModel chat.Chat, knowledge *types.Knowledge, chunks []*types.Chunk,
|
||||
) (string, error) {
|
||||
@@ -1919,6 +1922,12 @@ func (s *knowledgeService) getSummary(ctx context.Context,
|
||||
return "", fmt.Errorf("no chunks provided for summary generation")
|
||||
}
|
||||
|
||||
// Determine max input chars from config
|
||||
maxInputChars := defaultMaxInputChars
|
||||
if s.config.Conversation.Summary != nil && s.config.Conversation.Summary.MaxInputChars > 0 {
|
||||
maxInputChars = s.config.Conversation.Summary.MaxInputChars
|
||||
}
|
||||
|
||||
// concat chunk contents
|
||||
chunkContents := ""
|
||||
allImageInfos := make([]*types.ImageInfo, 0)
|
||||
@@ -1930,11 +1939,8 @@ func (s *knowledgeService) getSummary(ctx context.Context,
|
||||
return sortedChunks[i].StartAt < sortedChunks[j].StartAt
|
||||
})
|
||||
|
||||
// concat chunk contents and collect image infos
|
||||
// concat ALL chunk contents (no early truncation) and collect image infos
|
||||
for _, chunk := range sortedChunks {
|
||||
if chunk.EndAt > 4096 {
|
||||
break
|
||||
}
|
||||
// Ensure we don't slice beyond the current content length
|
||||
runes := []rune(chunkContents)
|
||||
if chunk.StartAt <= len(runes) {
|
||||
@@ -1970,9 +1976,11 @@ func (s *knowledgeService) getSummary(ctx context.Context,
|
||||
chunkContents = chunkContents + imageAnnotations
|
||||
}
|
||||
|
||||
if len(chunkContents) < 300 {
|
||||
return chunkContents, nil
|
||||
}
|
||||
// Apply length limit: sample long content to fit within maxInputChars
|
||||
chunkContents = sampleLongContent(chunkContents, maxInputChars)
|
||||
|
||||
logger.GetLogger(ctx).Infof("getSummary: content length=%d chars (max=%d) for knowledge %s",
|
||||
len([]rune(chunkContents)), maxInputChars, knowledge.ID)
|
||||
|
||||
// Prepare content with metadata for summary generation
|
||||
contentWithMetadata := chunkContents
|
||||
@@ -1990,6 +1998,12 @@ func (s *knowledgeService) getSummary(ctx context.Context,
|
||||
contentWithMetadata = metadataIntro + "\nContent:\n" + contentWithMetadata
|
||||
}
|
||||
|
||||
// Determine max output tokens from config
|
||||
maxTokens := 2048
|
||||
if s.config.Conversation.Summary != nil && s.config.Conversation.Summary.MaxCompletionTokens > 0 {
|
||||
maxTokens = s.config.Conversation.Summary.MaxCompletionTokens
|
||||
}
|
||||
|
||||
// Generate summary using AI model
|
||||
summaryPrompt := types.RenderPromptPlaceholders(s.config.Conversation.GenerateSummaryPrompt, types.PlaceholderValues{
|
||||
"language": types.LanguageNameFromContext(ctx),
|
||||
@@ -2006,7 +2020,7 @@ func (s *knowledgeService) getSummary(ctx context.Context,
|
||||
},
|
||||
}, &chat.ChatOptions{
|
||||
Temperature: 0.3,
|
||||
MaxTokens: 1024,
|
||||
MaxTokens: maxTokens,
|
||||
Thinking: &thinking,
|
||||
})
|
||||
if err != nil {
|
||||
@@ -2017,6 +2031,51 @@ func (s *knowledgeService) getSummary(ctx context.Context,
|
||||
return summary.Content, nil
|
||||
}
|
||||
|
||||
// sampleLongContent returns content that fits within maxChars.
|
||||
// For short content (≤ maxChars), it is returned as-is.
|
||||
// For long content, it samples: head (60%), tail (20%), and evenly-spaced middle (20%),
|
||||
// joined by "[...content omitted...]" markers so the LLM knows content was skipped.
|
||||
func sampleLongContent(content string, maxChars int) string {
|
||||
runes := []rune(content)
|
||||
if len(runes) <= maxChars {
|
||||
return content
|
||||
}
|
||||
|
||||
const omitMarker = "\n\n[...content omitted...]\n\n"
|
||||
omitRunes := len([]rune(omitMarker))
|
||||
|
||||
// Reserve space for two omit markers (head→middle, middle→tail)
|
||||
usable := maxChars - 2*omitRunes
|
||||
if usable < 100 {
|
||||
// Fallback: just truncate
|
||||
return string(runes[:maxChars])
|
||||
}
|
||||
|
||||
headLen := usable * 60 / 100
|
||||
tailLen := usable * 20 / 100
|
||||
midLen := usable - headLen - tailLen
|
||||
|
||||
head := string(runes[:headLen])
|
||||
tail := string(runes[len(runes)-tailLen:])
|
||||
|
||||
// Sample middle portion: take a contiguous block from the center of the document
|
||||
midStart := len(runes)/2 - midLen/2
|
||||
if midStart < headLen {
|
||||
midStart = headLen
|
||||
}
|
||||
midEnd := midStart + midLen
|
||||
if midEnd > len(runes)-tailLen {
|
||||
midEnd = len(runes) - tailLen
|
||||
midStart = midEnd - midLen
|
||||
if midStart < headLen {
|
||||
midStart = headLen
|
||||
}
|
||||
}
|
||||
middle := string(runes[midStart:midEnd])
|
||||
|
||||
return head + omitMarker + middle + omitMarker + tail
|
||||
}
|
||||
|
||||
// enqueueQuestionGenerationTask enqueues an async task for question generation
|
||||
func (s *knowledgeService) enqueueQuestionGenerationTask(ctx context.Context,
|
||||
kbID, knowledgeID string, questionCount int,
|
||||
|
||||
Reference in New Issue
Block a user