Chunking Strategy

Optimize how documents are split into memories for better retrieval and context preservation.

Chunking Methods

Fixed-Size Chunking

Split text into fixed-length chunks.

Best for: Simple documents, quick setup

om.ingest_file(
    'document.pdf',
    chunk_strategy='fixed',
    chunk_size=512,
    chunk_overlap=50
)

Pros:

• Fast and simple
• Predictable chunk sizes
• Low memory usage

Cons:

• May split sentences/paragraphs
• Loses semantic boundaries

Semantic Chunking

Split based on semantic coherence.

Best for: Articles, documentation, books

om.ingest_file(
    'article.md',
    chunk_strategy='semantic',
    similarity_threshold=0.75
)

Pros:

• Maintains topic coherence
• Better context preservation
• Improves retrieval quality

Cons:

• Slower processing
• Variable chunk sizes

Sentence-Based Chunking

Split at sentence boundaries.

Best for: Chat logs, Q&A, structured text

om.ingest_file(
    'conversation.txt',
    chunk_strategy='sentence',
    sentences_per_chunk=3
)

Code-Aware Chunking

Split code by functions/classes.

Best for: Source code repositories

om.ingest_file(
    'module.py',
    chunk_strategy='code',
    split_by='function'  # or 'class', 'method'
)

Configuration

Chunk Size Guidelines

Overlap Strategy

# High overlap - better context but more storage
om.ingest_file(
    'document.pdf',
    chunk_size=500,
    chunk_overlap=100  # 20% overlap
)

# Low overlap - less storage but may miss context
om.ingest_file(
    'document.pdf',
    chunk_size=500,
    chunk_overlap=25  # 5% overlap
)

Advanced Techniques

Hierarchical Chunking

# Create parent-child chunk relationships
om.ingest_file(
    'book.pdf',
    chunk_strategy='hierarchical',
    levels=[
        {'size': 2000, 'name': 'chapter'},
        {'size': 500, 'name': 'section'},
        {'size': 100, 'name': 'paragraph'}
    ]
)

Metadata-Enhanced Chunking

# Extract and add metadata to chunks
om.ingest_file(
    'document.pdf',
    extract_metadata=True,  # Headers, page numbers, etc.
    metadata_strategy='inherit'  # Inherit from document
)

Custom Chunking Function

def custom_chunker(text: str) -> list[str]:
    """Custom chunking logic"""
    chunks = []
    # Your logic here
    return chunks

om.ingest_file(
    'document.pdf',
    chunk_function=custom_chunker
)

Best Practices

1. Match chunk size to query length - Similar sizes work better
2. Use semantic chunking for quality - Worth the extra processing
3. Add overlap for context - 10-20% overlap recommended
4. Preserve structure - Keep paragraphs/sections together
5. Test and iterate - Evaluate retrieval quality

Performance Impact

See Multimodal Ingestion for file ingestion and Custom Providers for custom chunkers.