Chunking

class DocumentModelClass extends BaseModel {
  @upload({ mode: 'single' })
  @api({ httpMethod: 'POST' })
  async uploadDocument() {
    const { files } = Sonamu.getContext();
    const file = files?.[0]; // Use first file
    const content = await file.toBuffer().then(b => b.toString());

    // Attempting to generate embedding
    const embedding = await Embedding.embedOne(content, 'voyage', 'document');
    // Error: Token limit exceeded (32,000 tokens)
  }
}

import { Chunking } from "sonamu/vector";

const chunking = new Chunking({
  chunkSize: 500,        // chunk size (characters)
  chunkOverlap: 50,      // overlap size
  minChunkSize: 50,      // minimum size
  skipThreshold: 200,    // skip splitting if short
  separators: ['\n\n', '\n', '. '],  // separators
});

const chunks = chunking.chunk("Long text...");

class DocumentModelClass extends BaseModel {
  @upload({ mode: 'single' })
  @api({ httpMethod: 'POST' })
  async uploadLongDocument() {
    const { files } = Sonamu.getContext();
    const file = files?.[0]; // Use first file
    const content = await file.toBuffer().then(b => b.toString());

    // 1. Chunking
    const chunking = new Chunking({
      chunkSize: 500,
      chunkOverlap: 50,
    });

    const chunks = chunking.chunk(content);

    // 2. Embedding per chunk
    const embeddings = await Embedding.embed(
      chunks.map(c => c.text),
      'voyage',
      'document'
    );

    // 3. Create parent document
    const parent = await this.saveOne({
      title: file.filename,
      content,
      chunk_count: chunks.length,
    });

    // 4. Save each chunk
    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,
    };
  }
}

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

-- Chunks (store embeddings)
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,  // 500 characters
});

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

Chunk 1: [..............................]
Chunk 2:                    [...........]
                           ^ overlap area

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

// Don't split if 200 characters or less
const text = "Short text";  // 50 characters
const chunks = chunking.chunk(text);  // [entire text as 1 chunk]

const chunking = new Chunking({
  separators: [
    '\n\n',  // 1st priority: paragraph
    '\n',    // 2nd priority: line
    '. ',    // 3rd priority: sentence
    ', ',    // 4th priority: comma
  ],
});

@upload({ mode: 'single' })
@api({ httpMethod: 'POST' })
async uploadTechDoc() {
  const { files } = Sonamu.getContext();
  const file = files?.[0]; // Use first file
  const content = await file.toBuffer().then(b => b.toString());

  const chunking = new Chunking({
    chunkSize: 500,
    skipThreshold: 300,
  });

  // Keep as-is if short, chunk if long
  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) {
  // Same as the example above
  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');

  // Search chunks
  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,
  ]);

  // Group by parent document
  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": "Getting Started with TypeScript",
    "bestSimilarity": 0.89,
    "relevantChunks": [
      {
        "content": "TypeScript is JavaScript with types...",
        "chunkIndex": 2,
        "similarity": 0.89
      }
    ]
  }
]

const markdownChunking = new Chunking({
  chunkSize: 600,
  separators: [
    '\n## ',   // Heading 2
    '\n### ',  // Heading 3
    '\n\n',    // Paragraph
    '\n',      // Line
    '. ',      // Sentence
  ],
});

const markdown = `
# Sonamu

## Overview
Sonamu is a TypeScript framework.

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

const chunks = markdownChunking.chunk(markdown);

const chunking = new Chunking({
  skipThreshold: 300,  // Skip if 300 chars or less
});

if (chunking.needsChunking(content)) {
  // Process with chunking
} else {
  // Process as-is
}