DEV Community

vishalmysore
vishalmysore

Posted on

RAG Chunking Strategies Deep Dive

#systemdesign #rag #llm #ai

Retrieval-Augmented Generation (RAG) systems face a fundamental challenge: LLMs have context window limits, yet documents often exceed these limits. Simply stuffing an entire document into a prompt isn't feasible for large corpora. This is where chunking becomes critical.

Code for this article is https://github.com/vishalmysore/agenticmemory

The Chunking Challenge

Without proper chunking, RAG systems suffer from:

  • Lost context: Breaking text at arbitrary boundaries destroys semantic meaning
  • Poor retrieval: Overly large chunks reduce precision; overly small chunks lose context
  • Inefficient embedding: Vector databases work best with semantically coherent units
  • Token waste: Irrelevant information consumes precious context window space

Built-in chunking strategies solve these problems by providing intelligent, domain-aware text segmentation that preserves semantic boundaries while optimizing for retrieval performance.

What is Chunking?

Chunking is the process of breaking down large documents into smaller, semantically meaningful segments that can be:

  1. Embedded as dense vectors for similarity search
  2. Retrieved independently based on relevance to a query
  3. Fed to an LLM within its context window constraints

Effective chunking balances two competing goals:

  • Chunks must be small enough to be precise and fit within embedding model limits (typically 512-8192 tokens)
  • Chunks must be large enough to contain sufficient context for accurate retrieval and generation

The optimal chunking strategy depends on your document type, retrieval task, and downstream LLM usage.

Framework Overview

The Agentic Memory library includes an extensible chunking framework that allows you to split documents into optimal chunks for semantic search and retrieval.

Architecture

All chunking strategies are part of the core framework in the io.github.vishalmysore.rag.chunking package. The example code simply demonstrates how to use these strategies.

Core Components

  1. ChunkingStrategy interface - Base interface for all chunking strategies

    • List<String> chunk(String content) - Splits content into chunks
    • String getName() - Returns strategy name
    • String getDescription() - Returns strategy description

  2. RAGService.addDocumentWithChunking() - Convenience method for automatic chunking

   int chunkCount = rag.addDocumentWithChunking(
       "document_id", 
       content, 
       chunkingStrategy
   );
Enter fullscreen mode Exit fullscreen mode

Built-in Strategies

1. Sliding Window Chunking

Package: io.github.vishalmysore.rag.chunking.SlidingWindowChunking

Creates overlapping chunks to preserve context across boundaries.

Technical Details:

  • Uses a sliding window approach with configurable window size and overlap
  • Overlap percentage ensures context continuity between adjacent chunks
  • Word-based tokenization with configurable delimiters
  • Maintains approximately equal chunk sizes for consistent embedding quality 
ChunkingStrategy strategy = new SlidingWindowChunking(150, 30);
// 150 words per chunk, 30 words overlap (20%)
Enter fullscreen mode Exit fullscreen mode

Parameters:

  • windowSize: Number of words per chunk (typical: 100-300)
  • overlap: Number of overlapping words between chunks (typical: 10-20% of window size)

Best for: Healthcare records, continuous narratives, patient notes where context flows across boundaries

2. Adaptive Chunking

Package: io.github.vishalmysore.rag.chunking.AdaptiveChunking

Respects natural document boundaries while staying within token limits.

Technical Details:

  • Uses regex pattern matching to identify semantic boundaries (sections, paragraphs, etc.)
  • Dynamically adjusts chunk size based on boundary locations
  • Enforces min/max token constraints to balance precision and context
  • Preserves document structure by never splitting across matched boundaries 
ChunkingStrategy strategy = new AdaptiveChunking(
    "(?m)^SECTION \\d+:",  // Boundary pattern (regex)
    800,                    // Min tokens
    1200                    // Max tokens
);
Enter fullscreen mode Exit fullscreen mode

Parameters:

  • boundaryPattern: Regex to identify split points (e.g., section headers, paragraph markers)
  • minTokens: Minimum chunk size to maintain context
  • maxTokens: Maximum chunk size to fit embedding model limits

Best for: Legal contracts, structured documents, policy documents with clear section markers

3. Entity-Based Chunking

Package: io.github.vishalmysore.rag.chunking.EntityBasedChunking

Groups sentences by mentioned entities (people, companies, locations).

Technical Details:

  • Performs Named Entity Recognition (NER) on input text
  • Groups consecutive sentences that reference the same entities
  • Uses entity co-occurrence analysis to determine chunk boundaries
  • Maintains entity context within each chunk for improved retrieval precision 
String[] entities = {"Elon Musk", "Tesla", "SpaceX"};
ChunkingStrategy strategy = new EntityBasedChunking(entities);
Enter fullscreen mode Exit fullscreen mode

Parameters:

  • entities: Array of entity names to track (people, organizations, locations, etc.)
  • Optional: Entity types (PERSON, ORG, LOCATION) for automatic detection

Algorithm:

  1. Scan text for entity mentions
  2. Group sentences with shared entity references
  3. Create chunks when entity focus shifts
  4. Preserve co-occurrence relationships

Best for: News articles, research papers, multi-person biographies, documents with multiple actors

4. Topic/Theme-Based Chunking

Package: io.github.vishalmysore.rag.chunking.TopicBasedChunking

Groups content by underlying topics or themes.

Technical Details:

  • Uses topic modeling or keyword matching to identify thematic shifts
  • Regex-based topic boundary detection for structured documents
  • Optional: Latent Dirichlet Allocation (LDA) for unsupervised topic discovery
  • Creates semantically coherent chunks around single topics 
ChunkingStrategy strategy = new TopicBasedChunking(
    "(EDUCATION|CAREER|PATENTS):"
);
Enter fullscreen mode Exit fullscreen mode

Parameters:

  • topicPattern: Regex pattern to identify topic boundaries
  • Optional: Topic model configuration for unsupervised chunking

Best for: Research papers, technical documentation, structured content with explicit topic markers

5. Hybrid Chunking

Package: io.github.vishalmysore.rag.chunking.HybridChunking

Combines multiple strategies in a pipeline.

ChunkingStrategy adaptive = new AdaptiveChunking("(?m)^===\\s*$");
ChunkingStrategy topic = new TopicBasedChunking("(INTRO|BODY|CONCLUSION):");

ChunkingStrategy strategy = new HybridChunking(adaptive, topic);
Enter fullscreen mode Exit fullscreen mode

Best for: Complex documents requiring multi-stage processing

6. Task-Aware Chunking

Package: io.github.vishalmysore.rag.chunking.TaskAwareChunking

Adapts chunking based on downstream task (summarization/search/Q&A).

Technical Details:

  • Implements task-specific heuristics for optimal chunk sizing
  • Summarization: Small, focused chunks (50-100 tokens) for granular summaries
  • Search: Medium chunks (200-400 tokens) with function signatures + docstrings
  • Q&A: Large chunks (500-1000 tokens) preserving full context for accurate answers
  • Adjusts overlap and boundaries based on task requirements 
// For summarization (small chunks)
ChunkingStrategy strategy = new TaskAwareChunking(TaskType.SUMMARIZATION);

// For search (medium chunks - signatures + docstrings)
ChunkingStrategy strategy = new TaskAwareChunking(TaskType.SEARCH);

// For Q&A (large chunks - full context)
ChunkingStrategy strategy = new TaskAwareChunking(TaskType.QA);
Enter fullscreen mode Exit fullscreen mode

Task Types:

  • SUMMARIZATION: Optimized for generating summaries (small, atomic chunks)
  • SEARCH: Optimized for semantic search (medium chunks with metadata)
  • QA: Optimized for question answering (large, context-rich chunks)

Best for: Code repositories, multi-purpose document processing, systems serving multiple use cases

7. HTML Tag-Based Chunking

Package: io.github.vishalmysore.rag.chunking.HTMLTagBasedChunking

Uses HTML/XML structure to determine boundaries.

ChunkingStrategy strategy = new HTMLTagBasedChunking(
    "h2",   // Split on <h2> tags
    true    // Strip HTML tags from output
);
Enter fullscreen mode Exit fullscreen mode

Best for: Web content, HTML documentation, XML documents

8. Code-Specific Chunking

Package: io.github.vishalmysore.rag.chunking.CodeSpecificChunking

Uses language syntax for AST-inspired boundaries.

Technical Details:

  • Parses code into Abstract Syntax Tree (AST) nodes
  • Chunks by logical units: classes, functions, methods, modules
  • Preserves code structure and dependencies
  • Maintains indentation and scope boundaries
  • Supports Python, Java, JavaScript with language-specific parsers 
ChunkingStrategy strategy = new CodeSpecificChunking(
    CodeSpecificChunking.Language.PYTHON
);
// Also supports JAVA and JAVASCRIPT
Enter fullscreen mode Exit fullscreen mode

Chunking Units:

  • Classes: Complete class definitions with methods
  • Functions/Methods: Individual functions with docstrings
  • Modules: Top-level module code
  • Imports: Grouped import statements

Best for: Source code repositories, code documentation, programming tutorials, code search systems

9. Regex Chunking

Package: io.github.vishalmysore.rag.chunking.RegexChunking

Uses regex patterns to identify boundaries and group chunks.

Technical Details:

  • Dual-pattern approach: split pattern + optional grouping pattern
  • Split pattern defines chunk boundaries
  • Grouping pattern clusters related chunks together
  • Supports complex regex with lookaheads, capture groups, and backreferences
  • Ideal for highly structured or predictable text formats 
ChunkingStrategy strategy = new RegexChunking(
    "\\[\\d{4}-\\d{2}-\\d{2} \\d{2}:\\d{2}:\\d{2}\\]",  // Split pattern
    "(ERROR|WARN|INFO|DEBUG)"                            // Group pattern
);
Enter fullscreen mode Exit fullscreen mode

Parameters:

  • splitPattern: Regex to identify chunk boundaries (e.g., timestamps, delimiters)
  • groupPattern: Optional regex to group related chunks by common attributes (e.g., log levels)

Use Cases:

  • Log files with timestamps
  • Structured data with consistent delimiters
  • Documents with predictable formatting

Best for: Server logs, structured text, timestamp-based data, CSV-like formats

Creating Custom Strategies

The framework is fully extensible. Create your own chunking strategy by implementing ChunkingStrategy:

public class MyCustomChunking implements ChunkingStrategy {

    @Override
    public List<String> chunk(String content) {
        // Your chunking logic here
        List<String> chunks = new ArrayList<>();
        // ... split content into chunks ...
        return chunks;
    }

    @Override
    public String getName() {
        return "My Custom Chunking";
    }

    @Override
    public String getDescription() {
        return "Splits content using my custom algorithm";
    }
}
Enter fullscreen mode Exit fullscreen mode

Then use it with RAGService:

ChunkingStrategy myStrategy = new MyCustomChunking();
rag.addDocumentWithChunking("doc_id", content, myStrategy);
Enter fullscreen mode Exit fullscreen mode

See CustomChunkingExample.java for a complete example.

Usage Examples

Basic Usage 

// 1. Create a chunking strategy
ChunkingStrategy strategy = new SlidingWindowChunking(150, 30);

// 2. Use it with RAGService
try (RAGService rag = new RAGService(indexPath, embeddingProvider)) {
    int chunkCount = rag.addDocumentWithChunking(
        "my_document",
        documentContent,
        strategy
    );
    rag.commit();

    System.out.println("Created " + chunkCount + " chunks");
}
Enter fullscreen mode Exit fullscreen mode

Manual Chunking

If you prefer to manually control chunk IDs:

ChunkingStrategy strategy = new EntityBasedChunking("Entity1", "Entity2");
List<String> chunks = strategy.chunk(content);

for (int i = 0; i < chunks.size(); i++) {
    String customId = "custom_prefix_" + i;
    rag.addDocument(customId, chunks.get(i));
}
Enter fullscreen mode Exit fullscreen mode

Running the Examples

Comprehensive Demo (All 9 Strategies) 

export OPENAI_API_KEY="your-key-here"
mvn exec:java -Dexec.mainClass="io.github.vishalmysore.rag.examples.ChunkingStrategiesExample" -q
Enter fullscreen mode Exit fullscreen mode

Custom Strategy Example 

export OPENAI_API_KEY="your-key-here"
mvn exec:java -Dexec.mainClass="io.github.vishalmysore.rag.examples.CustomChunkingExample" -q
Enter fullscreen mode Exit fullscreen mode

Design Principles

  1. Core Framework Integration - All strategies are part of the core library, not examples
  2. Extensibility - Simple interface allows custom implementations
  3. Composability - Strategies can be combined (see HybridChunking)
  4. Task-Specific - Different strategies for different use cases
  5. Zero Configuration - Sensible defaults with optional customization
  6. Semantic Preservation - Chunks maintain meaningful boundaries, not arbitrary splits
  7. Performance Optimized - Efficient algorithms for large-scale document processing

Technical Considerations

Chunk Size Guidelines

Embedding Model Recommended Chunk Size Max Tokens
OpenAI text-embedding-ada-002 200-500 words 8191
OpenAI text-embedding-3-small 200-500 words 8191
Sentence Transformers 100-300 words 512
Cohere embed-english-v3.0 300-600 words 512

Overlap Strategies

Why Overlap Matters:

  • Prevents information loss at chunk boundaries
  • Improves retrieval recall for queries spanning boundaries
  • Typical overlap: 10-20% of chunk size

Trade-offs:

  • Higher overlap = Better recall, more storage, slower indexing
  • Lower overlap = Faster indexing, less storage, potential information loss

Performance Metrics

Evaluate chunking quality by:

  1. Retrieval Precision: Do chunks contain relevant information?
  2. Context Completeness: Do chunks have enough context for the LLM?
  3. Semantic Coherence: Are chunks semantically self-contained?
  4. Boundary Quality: Do breaks occur at natural boundaries?

When to Use Which Strategy

Use Case Recommended Strategy
Healthcare records, patient notes Sliding Window
Legal contracts, policies Adaptive
News articles, multi-entity docs Entity-Based
Research papers, structured content Topic-Based
Complex multi-format docs Hybrid
Code repositories Code-Specific or Task-Aware
Web pages, HTML docs HTML Tag-Based
Server logs, timestamped data Regex

API Reference

ChunkingStrategy Interface 

public interface ChunkingStrategy {
    List<String> chunk(String content);
    String getName();
    String getDescription();
}
Enter fullscreen mode Exit fullscreen mode

RAGService Methods 

// Add document with automatic chunking
int addDocumentWithChunking(String baseId, String content, ChunkingStrategy strategy)

// Add document with chunking and metadata
int addDocumentWithChunking(String baseId, String content, ChunkingStrategy strategy, 
                           Map<String, String> metadata)
Enter fullscreen mode Exit fullscreen mode

Advanced Topics

Chunking for Multi-Modal Documents

When working with documents containing mixed content (text, tables, images):

  1. Use HTML Tag-Based Chunking to separate content types
  2. Apply different strategies per content type
  3. Maintain cross-references between chunks

Chunking at Scale

For processing millions of documents:

  • Batch processing: Use parallel streams for chunking
  • Incremental indexing: Chunk and index in batches
  • Memory management: Process large documents in streaming mode
  • Caching: Cache compiled regex patterns and AST parsers

Dynamic Strategy Selection

Choose strategies programmatically based on document characteristics:

ChunkingStrategy selectStrategy(Document doc) {
    if (doc.isCode()) return new CodeSpecificChunking(detectLanguage(doc));
    if (doc.hasStructure()) return new AdaptiveChunking(detectBoundaries(doc));
    if (doc.hasEntities()) return new EntityBasedChunking(extractEntities(doc));
    return new SlidingWindowChunking(200, 40); // fallback
}
Enter fullscreen mode Exit fullscreen mode

Top comments (0)