llm-splitter

Splits text into semantically coherent chunks by respecting natural language boundaries (sentences, paragraphs, sections) rather than naive character/token limits. Implements configurable splitting strategies that preserve context integrity across chunk boundaries, enabling downstream LLM vectorization to capture meaningful semantic units. The chunker analyzes text structure and applies rule-based or learned boundary detection to minimize context fragmentation.

Automatically generates and attaches rich metadata to each chunk including byte/character offsets, chunk indices, original document position, and boundary type information. This metadata enables downstream systems to reconstruct document context, trace embeddings back to source locations, and implement overlap-aware retrieval strategies. The implementation tracks position state throughout the splitting process to ensure accurate offset calculation.

Exposes configuration parameters for chunk size (in characters or tokens), overlap amount, and splitting strategy selection, allowing users to tune chunking behavior for specific use cases without code changes. Implements parameter validation and applies configurations consistently across the splitting pipeline. Supports both fixed-size and adaptive sizing strategies based on document structure.

Implements multiple splitting strategies (recursive character splitting, sentence-aware splitting, paragraph-aware splitting) that can be selected or composed based on document type and requirements. Each strategy applies different boundary detection heuristics (punctuation, whitespace, structural markers) to identify natural break points. The implementation allows strategy composition to handle mixed-format documents.

Optimizes chunking performance for large-scale document processing by implementing efficient batch operations and minimal memory overhead. The implementation processes text sequentially with streaming-friendly patterns, avoiding full document loading into memory. Designed specifically for integration into vectorization pipelines where throughput and memory efficiency are critical.

Detects natural text boundaries (sentence ends, paragraph breaks, section headers) using language-agnostic heuristics based on punctuation, whitespace, and structural patterns rather than language-specific NLP models. Applies rule-based detection across multiple languages without requiring language identification or language-specific models. Boundary detection is configurable to handle domain-specific patterns.