Skip to main content
Node parsers transform documents into smaller chunks (nodes) that are optimized for embedding and retrieval. Effective chunking is critical for RAG performance.

​
Why Chunking Matters

Chunking breaks large documents into smaller pieces because:
  • Embedding models have token limits - Most models work best with 512-2048 tokens
  • Better semantic granularity - Smaller chunks provide more precise retrieval
  • Improved context relevance - Return only the most relevant sections to the LLM
  • Efficient processing - Easier to embed and index smaller text segments

​
SentenceSplitter

The most commonly used parser that splits text while respecting sentence boundaries.

​
Basic Usage

import { SentenceSplitter, Document } from "llamaindex";

const splitter = new SentenceSplitter({
  chunkSize: 1024,
  chunkOverlap: 20
});

const document = new Document({
  text: "Your long document text here..."
});

const nodes = await splitter.transform([document]);

​
Configuration Options

const splitter = new SentenceSplitter({
  // Maximum tokens per chunk
  chunkSize: 1024,
  
  // Overlap between chunks (in tokens)
  chunkOverlap: 200,
  
  // Separator between paragraphs
  paragraphSeparator: "\n\n\n",
  
  // Secondary chunking regex for fallback
  secondaryChunkingRegex: "[^,.;。？！]+[,.;。？！]?",
  
  // Separator for splitting into words
  separator: " ",
  
  // Additional abbreviations to recognize (e.g., "LLC.")
  extraAbbreviations: ["LLC", "Inc"]
});

​
How It Works

  1. Paragraph splitting: First tries to split by paragraph separators
  2. Sentence splitting: Uses sentence tokenizer to find sentence boundaries
  3. Regex fallback: If sentences are too long, uses secondary regex
  4. Word splitting: Final fallback splits by words
  5. Chunk merging: Combines splits into chunks up to chunkSize with chunkOverlap

​
Metadata-Aware Splitting

const document = new Document({
  text: "Content here",
  metadata: {
    title: "Long Document Title Here",
    author: "Author Name"
  }
});

const splitter = new SentenceSplitter({ chunkSize: 1024 });
const nodes = await splitter.transform([document]);

// Effective chunk size is reduced by metadata length
// to ensure total content fits within chunkSize

​
MarkdownNodeParser

Splits markdown documents by headers, preserving document structure. 
import { MarkdownNodeParser, Document } from "llamaindex";

const markdown = `
# Main Title

Introduction text here.

## Section 1

Content for section 1.

### Subsection 1.1

Detailed content.

## Section 2

Content for section 2.
`;

const parser = new MarkdownNodeParser();
const document = new Document({ text: markdown });
const nodes = await parser.transform([document]);

// Each node contains text from one section
// Metadata includes header hierarchy:
// { Header_1: "Main Title", Header_2: "Section 1", Header_3: "Subsection 1.1" }

​
Features

  • Splits on markdown headers (#, ##, ###, etc.)
  • Preserves header hierarchy in metadata
  • Handles code blocks correctly
  • Each chunk contains one section’s content

​
CodeSplitter

Parses code using tree-sitter for syntax-aware chunking. 
import { CodeSplitter } from "@llamaindex/node-parser/code";
import Parser from "tree-sitter";
import TypeScript from "tree-sitter-typescript";

const parser = new Parser();
parser.setLanguage(TypeScript.typescript);

const codeSplitter = new CodeSplitter({
  getParser: () => parser,
  maxChars: 1500
});

const codeDocument = new Document({
  text: `
    export function example() {
      // Your code here
    }
    
    export class MyClass {
      // Class implementation
    }
  `
});

const nodes = await codeSplitter.transform([codeDocument]);

​
Features

  • Syntax-aware: Respects language structure (functions, classes, etc.)
  • Configurable size: Set maxChars for chunk length
  • Multi-language: Works with any tree-sitter grammar
  • Recursive chunking: Splits large syntax nodes intelligently

​
SentenceWindowNodeParser

Creates overlapping windows around sentences for better context. 
import { SentenceWindowNodeParser, Document } from "llamaindex";

const parser = new SentenceWindowNodeParser({
  windowSize: 3,  // 3 sentences before and after
  windowMetadataKey: "window",
  originalTextMetadataKey: "original_sentence"
});

const nodes = await parser.transform([document]);

// Each node contains one sentence with surrounding context
// Useful for more precise retrieval with expanded context

​
TokenTextSplitter

Splits text by token count without respecting sentence boundaries. 
import { TokenTextSplitter } from "llamaindex";

const splitter = new TokenTextSplitter({
  chunkSize: 512,
  chunkOverlap: 50,
  separator: " "
});

const nodes = await splitter.transform([document]);

​
SimpleNodeParser (Deprecated)

SimpleNodeParser is deprecated. Use SentenceSplitter instead.
// Old way (deprecated)
import { SimpleNodeParser } from "llamaindex";

// New way
import { SentenceSplitter } from "llamaindex";
const parser = new SentenceSplitter();

​
Custom Parsers

Create your own parser by extending NodeParser: 
import { NodeParser, TextNode, Document } from "llamaindex";

class CustomParser extends NodeParser {
  protected parseNodes(documents: TextNode[]): TextNode[] {
    // Your custom parsing logic
    const nodes: TextNode[] = [];
    
    for (const doc of documents) {
      const text = doc.getContent(MetadataMode.NONE);
      
      // Split by your custom logic
      const chunks = this.customSplit(text);
      
      // Create nodes from chunks
      for (const chunk of chunks) {
        const node = new TextNode({
          text: chunk,
          metadata: { ...doc.metadata }
        });
        node.relationships[NodeRelationship.SOURCE] = 
          doc.asRelatedNodeInfo();
        nodes.push(node);
      }
    }
    
    return nodes;
  }
  
  private customSplit(text: string): string[] {
    // Your splitting logic here
    return text.split("\n\n");
  }
}

const parser = new CustomParser();
const nodes = await parser.transform([document]);

​
Choosing a Chunking Strategy

Use SentenceSplitter with:
  • chunkSize: 1024 for most cases
  • chunkSize: 512 for more precise retrieval
  • chunkSize: 2048 for broader context
  • chunkOverlap: 200 to maintain continuity
Use MarkdownNodeParser to:
  • Preserve document structure
  • Keep sections together
  • Add header hierarchy to metadata
  • Improve navigation and citations
Use CodeSplitter to:
  • Respect syntax boundaries
  • Keep functions/classes intact
  • Enable code search and analysis
  • Support multiple languages
Use SentenceWindowNodeParser to:
  • Retrieve exact sentences
  • Provide surrounding context
  • Improve answer accuracy
  • Support citation to specific sentences

​
Complete Example

import { 
  Document, 
  SentenceSplitter,
  VectorStoreIndex,
  MetadataMode
} from "llamaindex";
import { OpenAIEmbedding } from "@llamaindex/openai";
import fs from "fs/promises";

async function main() {
  // Load document
  const text = await fs.readFile("article.txt", "utf-8");
  
  const document = new Document({
    text,
    metadata: {
      source: "article.txt",
      category: "technical"
    }
  });
  
  // Configure parser
  const parser = new SentenceSplitter({
    chunkSize: 1024,
    chunkOverlap: 200
  });
  
  // Split into nodes
  const nodes = await parser.transform([document]);
  
  console.log(`Created ${nodes.length} nodes`);
  console.log("First node:", nodes[0].text);
  console.log("Node metadata:", nodes[0].metadata);
  
  // Build index from nodes
  const index = await VectorStoreIndex.init({ 
    nodes,
    embedModel: new OpenAIEmbedding()
  });
  
  // Query
  const queryEngine = index.asQueryEngine();
  const response = await queryEngine.query({
    query: "What is this document about?"
  });
  
  console.log(response.toString());
}

main().catch(console.error);

​
Best Practices

  1. Match chunk size to your use case
    • Smaller (512) for precise retrieval
    • Larger (2048) for broad context
  2. Use appropriate overlap
    • 10-20% of chunk size typically works well
    • Prevents losing context at boundaries
  3. Respect document structure
    • Use MarkdownNodeParser for markdown
    • Use CodeSplitter for code
    • Don’t split across major boundaries
  4. Consider metadata
    • Account for metadata in chunk size
    • Use metadata to preserve structure
    • Add custom fields for filtering
  5. Test your strategy
    • Evaluate retrieval quality
    • Adjust chunk size based on results
    • Monitor token usage

​
Next Steps

Documents

Learn about Document structure

Ingestion

Build complete processing pipelines

Embeddings

Configure embedding models

Retrieval

Optimize retrieval strategies

Build docs developers (and LLMs) love

Get started for freeTalk to us