summary

AI Summarizer (ai/summary)

The summary package provides intelligent summary generation service for quickly extracting the core content of Memos.

Architecture

classDiagram
    class Summarizer {
        <<interface>>
        +Summarize(ctx, req)
    }
    class llmSummarizer {
        -llm Service
        -timeout Duration
        +Summarize()
    }
    class FallbackSummarizer {
        +Summarize()
        -truncate()
        -firstParagraph()
    }

    Summarizer <|.. llmSummarizer
    llmSummarizer --> FallbackSummarizer : uses on fail

• Summarizer Interface: Defines summary generation capability.
• Fallback Strategy: Considering LLM may timeout or be unavailable, the system implements a multi-level fallback strategy.

Algorithm Flow

flowchart TD
    Start[Request] --> ShortText{Content <= MaxLen?}
    ShortText -- Yes --> ReturnOriginal[Return Original Content]
    ShortText -- No --> TryLLM{LLM Available?}
    TryLLM -- Success --> ReturnLLM[Return LLM Summary]
    TryLLM -- Fail/Timeout --> Fallback{Fallback Levels}

    Fallback --> Level1[First Paragraph]
    Level1 -- Too Short --> Level2[First Sentence]
    Level2 -- Too Short --> Level3[Truncate 50 chars]

    Level3 --> ReturnFallback[Return Fallback Summary]

Algorithm Details

1. Short Text Optimization

If text length (in runes) <= MaxLen (default 200), return original content directly with Source: original.

2. LLM Generation

Use a dedicated prompt (e.g., "请用一句话概括...") to call LLM.

• Request LLM to return JSON format: {"summary": "生成的摘要"}
• Parse and truncate to MaxLen characters.

3. Automatic Fallback

If LLM call fails, times out, or returns empty, try in order:

• First Paragraph: Extract first paragraph.
• First Sentence: Extract first sentence.
• Truncate: Truncate first N characters (default 50).

4. Result Marking

Return result includes Source field (llm, fallback_*, original) for frontend to distinguish display or decide whether to retry.

Configuration