Struct LateChunkingStrategy 

pub struct LateChunkingStrategy { /* private fields */ }

Context-aware chunking strategy for use with late-chunking embedding models

Splits text using HierarchicalChunker and records precise byte-offset spans in each chunk’s metadata. A late-chunking embedding provider (Jina API or a local candle model) can then use these spans to extract per-chunk representations from a single full-document forward pass.

Examples

use graphrag_core::text::late_chunking::{LateChunkingStrategy, LateChunkingConfig};
use graphrag_core::core::{ChunkingStrategy, DocumentId};

let strategy = LateChunkingStrategy::with_defaults(DocumentId::new("doc-1".to_string()));
let chunks = strategy.chunk("First paragraph.\n\nSecond paragraph.");

for chunk in &chunks {
    // position_in_document ∈ [0.0, 1.0] — used by embedding provider for pooling
    assert!(chunk.metadata.position_in_document.is_some());
}

Implementations

impl LateChunkingStrategy

pub fn new(config: LateChunkingConfig, document_id: DocumentId) -> Self

Create a new late chunking strategy with explicit config

pub fn with_defaults(document_id: DocumentId) -> Self

Create with default config (8192 token limit, 512-char chunks)

pub fn with_max_doc_tokens(self, max_tokens: u32) -> Self

Set the maximum document token limit (choose based on embedding model)

• 8192 → Jina v3
• 32768 → gte-Qwen2-7B-instruct

pub fn estimate_tokens(text: &str) -> u32

Estimate token count from character count (1 token ≈ 4 chars)

pub fn fits_in_context(&self, text: &str) -> bool

Returns true if the document fits within the model’s context window

pub fn split_into_sections(&self, text: &str) -> Vec<String>

Split an oversized document into sections that fit within the context window

Sections are formed by grouping paragraphs (double-newline boundaries) until the next paragraph would exceed the limit. Each section can be embedded independently with late chunking applied within it.

Trait Implementations

impl ChunkingStrategy for LateChunkingStrategy

fn chunk(&self, text: &str) -> Vec<TextChunk>

Chunk text into pieces following the strategy’s logic