청킹

class DocumentModelClass extends BaseModelClass {
  @upload()
  async uploadDocument() {
    const { bufferedFiles } = Sonamu.getContext();
    const file = bufferedFiles?.[0]; // 첫 번째 파일 사용
    const content = file.buffer.toString();

    // 임베딩 생성 시도
    const embedding = await Embedding.embedOne(content, 'voyage', 'document');
    // ❌ 에러: 토큰 제한 초과 (32,000 토큰)
  }
}

import { Chunking } from "sonamu/vector";

const chunking = new Chunking({
  chunkSize: 500,        // 청크 크기 (문자)
  chunkOverlap: 50,      // 중복 크기
  minChunkSize: 50,      // 최소 크기
  skipThreshold: 200,    // 짧으면 분할 스킵
  separators: ['\n\n', '\n', '. '],  // 구분자
});

const chunks = chunking.chunk("긴 텍스트...");

class DocumentModelClass extends BaseModelClass {
  @upload()
  async uploadLongDocument() {
    const { bufferedFiles } = Sonamu.getContext();
    const file = bufferedFiles?.[0]; // 첫 번째 파일 사용
    const content = file.buffer.toString();
    
    // 1. 청킹
    const chunking = new Chunking({
      chunkSize: 500,
      chunkOverlap: 50,
    });
    
    const chunks = chunking.chunk(content);
    
    // 2. 각 청크별 임베딩
    const embeddings = await Embedding.embed(
      chunks.map(c => c.text),
      'voyage',
      'document'
    );
    
    // 3. 부모 문서 생성
    const parent = await this.saveOne({
      title: file.filename,
      content,
      chunk_count: chunks.length,
    });
    
    // 4. 청크별 저장
    const savedChunks = await Promise.all(
      chunks.map((chunk, i) => 
        DocumentChunkModel.saveOne({
          parent_id: parent.id,
          chunk_index: chunk.index,
          content: chunk.text,
          start_offset: chunk.startOffset,
          end_offset: chunk.endOffset,
          embedding: embeddings[i].embedding,
        })
      )
    );
    
    return {
      parentId: parent.id,
      chunkCount: chunks.length,
    };
  }
}

-- 부모 문서
CREATE TABLE documents (
  id SERIAL PRIMARY KEY,
  title TEXT NOT NULL,
  content TEXT NOT NULL,
  chunk_count INTEGER,
  created_at TIMESTAMP DEFAULT NOW()
);

-- 청크 (임베딩 저장)
CREATE TABLE document_chunks (
  id SERIAL PRIMARY KEY,
  parent_id INTEGER REFERENCES documents(id),
  chunk_index INTEGER,
  content TEXT NOT NULL,
  start_offset INTEGER,
  end_offset INTEGER,
  embedding vector(1024),
  created_at TIMESTAMP DEFAULT NOW()
);

CREATE INDEX ON document_chunks (parent_id);
CREATE INDEX ON document_chunks USING hnsw (embedding vector_cosine_ops);

const chunking = new Chunking({
  chunkSize: 500,  // 500자
});

const chunking = new Chunking({
  chunkSize: 500,
  chunkOverlap: 50,  // 10%
});

청크 1: [..............................]
청크 2:                    [...........]
                           ↑ 중복 영역

const chunking = new Chunking({
  skipThreshold: 200,
});

// 200자 이하는 분할 안 함
const text = "짧은 글";  // 50자
const chunks = chunking.chunk(text);  // [전체 텍스트 1개]

const chunking = new Chunking({
  separators: [
    '\n\n',  // 1순위: 단락
    '\n',    // 2순위: 줄
    '. ',    // 3순위: 문장
    ', ',    // 4순위: 쉼표
  ],
});

@upload()
async uploadTechDoc() {
  const { bufferedFiles } = Sonamu.getContext();
  const file = bufferedFiles?.[0]; // 첫 번째 파일 사용
  const content = file.buffer.toString();
  
  const chunking = new Chunking({
    chunkSize: 500,
    skipThreshold: 300,
  });
  
  // 짧으면 그대로, 길면 청킹
  if (chunking.needsChunking(content)) {
    return await this.uploadWithChunking(file.filename, content);
  } else {
    return await this.uploadSimple(file.filename, content);
  }
}

private async uploadSimple(title: string, content: string) {
  const embedding = await Embedding.embedOne(
    `${title}\n\n${content}`,
    'voyage',
    'document'
  );
  
  return await this.saveOne({
    title,
    content,
    embedding: embedding.embedding,
  });
}

private async uploadWithChunking(title: string, content: string) {
  // 위의 예제와 동일
  const chunking = new Chunking({ chunkSize: 500 });
  const chunks = chunking.chunk(content);
  // ...
}

@api({ httpMethod: 'POST' })
async searchDocs(query: string, limit: number = 5) {
  const embedding = await Embedding.embedOne(query, 'voyage', 'query');
  
  // 청크 검색
  const chunks = await this.getPuri().raw(`
    SELECT 
      c.id, c.parent_id, c.content, c.chunk_index,
      d.title,
      1 - (c.embedding <=> ?) AS similarity
    FROM document_chunks c
    JOIN documents d ON c.parent_id = d.id
    WHERE c.embedding IS NOT NULL
    ORDER BY c.embedding <=> ?
    LIMIT ?
  `, [
    JSON.stringify(embedding.embedding),
    JSON.stringify(embedding.embedding),
    limit * 2,
  ]);
  
  // 부모 문서별 그룹화
  const grouped = new Map();
  
  for (const chunk of chunks.rows) {
    const parentId = chunk.parent_id;
    
    if (!grouped.has(parentId)) {
      grouped.set(parentId, {
        parentId,
        title: chunk.title,
        bestSimilarity: chunk.similarity,
        relevantChunks: [],
      });
    }
    
    grouped.get(parentId).relevantChunks.push({
      content: chunk.content,
      chunkIndex: chunk.chunk_index,
      similarity: chunk.similarity,
    });
  }
  
  return Array.from(grouped.values())
    .sort((a, b) => b.bestSimilarity - a.bestSimilarity)
    .slice(0, limit);
}

[
  {
    "parentId": 123,
    "title": "TypeScript 시작하기",
    "bestSimilarity": 0.89,
    "relevantChunks": [
      {
        "content": "TypeScript는 타입이 있는 JavaScript입니다...",
        "chunkIndex": 2,
        "similarity": 0.89
      }
    ]
  }
]

const markdownChunking = new Chunking({
  chunkSize: 600,
  separators: [
    '\n## ',   // 헤딩 2
    '\n### ',  // 헤딩 3
    '\n\n',    // 단락
    '\n',      // 줄
    '. ',      // 문장
  ],
});

const markdown = `
# Sonamu

## 개요
Sonamu는 TypeScript 프레임워크입니다.

## 설치
\`\`\`bash
pnpm add sonamu
\`\`\`
`;

const chunks = markdownChunking.chunk(markdown);

const chunking = new Chunking({
  skipThreshold: 300,  // 300자 이하는 스킵
});

if (chunking.needsChunking(content)) {
  // 청킹 처리
} else {
  // 그대로 처리
}