Chonkie 1 6 2

Chonkie 1.6.2

Overview

Chonkie is a lightweight, high-performance text chunking library designed for retrieval-augmented generation (RAG) pipelines. It provides 12 chunking strategies ranging from simple token-based splitting to neural and agentic approaches, all unified behind a consistent interface. Built with Python (3.10+) and JavaScript support, Chonkie follows the CHOMP architecture — Chef (preprocess), Hchunk, Overlap/Refine, Merge/export, Port/store — to orchestrate end-to-end document ingestion workflows.

Key characteristics:

• Lightweight: 505 KB wheel size vs 1–12 MB for alternatives. Base install pulls only ~4 core dependencies (tqdm, numpy, chonkie-core, tenacity).
• Fast: SIMD-accelerated chunking via chonkie-core Rust extension. Benchmarks show 33x faster token chunking than the slowest competitor and 2x faster sentence chunking than LlamaIndex.
• Modular dependencies: Install only what you need via optional extras ([semantic], [code], [neural], etc.).
• Unified Chunk type: Since v1.3.0, all chunkers return the same base Chunk dataclass with fields: text, start_index, end_index, token_count, optional context, and optional embedding.
• Async support: Every chunker provides achunk(), achunk_batch(), and achunk_document() methods out of the box using asyncio.to_thread.
• 32+ integrations: Tokenizers (tiktoken, HuggingFace tokenizers), embedding providers (OpenAI, Cohere, Jina, Voyage AI, sentence-transformers, Model2Vec, LiteLLM), LLM genies (Gemini, OpenAI, Groq, Cerebras), and vector database handshakes (ChromaDB, Qdrant, Pinecone, Weaviate, pgvector, MongoDB, Elasticsearch, Milvus, Turbopuffer).

When to Use

• Building RAG pipelines that need high-quality text chunking
• Splitting documents into semantically coherent chunks for vector database ingestion
• Processing code files with structure-aware chunking (AST-based CodeChunker)
• Running end-to-end ingestion workflows: fetch → clean → chunk → refine → embed → store
• Needing fast, lightweight chunking without the overhead of LangChain or LlamaIndex
• Building self-hosted REST APIs for chunking services via chonkie serve

Core Concepts

Chunk: The fundamental output unit. A dataclass with text, start_index, end_index, token_count, optional context, and optional embedding. All chunkers return this same type.

Chunker: A component that splits text into Chunk objects. Chonkie provides 12 chunkers: Token, Fast, Sentence, Recursive, Semantic, Late, Code, Neural, Slumber, Table, TeraflopAI, and SDPM (legacy).

Refinery: A post-processing component that enhances chunks. OverlapRefinery adds context from neighboring chunks. EmbeddingsRefinery computes and attaches embedding vectors.

Pipeline: A fluent, chainable interface for building multi-step workflows following the CHOMP architecture: Fetcher → Chef → Chunker → Refinery → Porter/Handshake. Pipelines auto-reorder components into correct execution order.

Chef: Text preprocessing component. TextChef cleans and normalizes text. MarkdownChef extracts tables and code blocks from markdown.

Porter: Exports chunks to file formats. JSONPorter writes JSON. DatasetsPorter pushes to HuggingFace Datasets.

Handshake: Connects chunks directly to vector databases for embedding and storage in one step (ChromaDB, Qdrant, Pinecone, Weaviate, pgvector, MongoDB, Elasticsearch, Milvus, Turbopuffer).

Genie: Interface to LLM providers for advanced chunking strategies. Supports Gemini, OpenAI, Azure OpenAI, Groq, and Cerebras.

Recipe: Pre-configured chunker settings loaded from HuggingFace Hub. Use from_recipe() for language-specific or document-type-specific chunking (e.g., RecursiveChunker.from_recipe("markdown", lang="en")).

Installation / Setup

Basic Python Installation

pip install chonkie

Or with uv:

uv add chonkie

This provides TokenChunker, FastChunker, SentenceChunker, RecursiveChunker, TableChunker, OverlapRefinery, and basic tokenizers (character, word, byte).

Optional Features

Install specific capabilities as needed:

# Semantic chunking (SemanticChunker, LateChunker) with Model2Vec embeddings
pip install "chonkie[semantic]"

# Code-aware chunking (CodeChunker) with tree-sitter
pip install "chonkie[code]"

# Neural chunking (NeuralChunker) with transformers + torch
pip install "chonkie[neural]"

# LLM-based chunking (SlumberChunker) via Genie interface
pip install "chonkie[genie]"

# Visualization tools
pip install "chonkie[viz]"

# HuggingFace Hub recipes
pip install "chonkie[hub]"

# All features (not recommended for production)
pip install "chonkie[all]"

# Multiple features combined
pip install "chonkie[semantic,code,viz]"

JavaScript Installation

npm install @chonkiejs/core     # Local chunking (Token + Recursive)
npm install @chonkiejs/cloud    # API client for cloud chunking
npm install @chonkiejs/token    # Custom tokenizers for JS

Logging Control

export CHONKIE_LOG=off      # Disable all logging
export CHONKIE_LOG=warning  # Warnings and errors (default)
export CHONKIE_LOG=info     # More verbose
export CHONKIE_LOG=debug    # Everything

Usage Examples

Basic Chunking

from chonkie import RecursiveChunker

chunker = RecursiveChunker(chunk_size=512)
chunks = chunker("Your document text here...")

for chunk in chunks:
    print(f"Text: {chunk.text[:50]}...")
    print(f"Tokens: {chunk.token_count}")

Pipeline Workflow

from chonkie import Pipeline

doc = (Pipeline()
    .chunk_with("recursive", chunk_size=512)
    .refine_with("overlap", context_size=100)
    .run(texts="Your document text here..."))

for chunk in doc.chunks:
    print(chunk.text)

RAG Ingestion Pipeline

from chonkie import Pipeline

docs = (Pipeline()
    .fetch_from("file", dir="./knowledge_base", ext=[".txt", ".md"])
    .process_with("text")
    .chunk_with("semantic", threshold=0.8, chunk_size=1024)
    .refine_with("overlap", context_size=100)
    .store_in("qdrant", collection_name="knowledge", url="http://localhost:6333")
    .run())

Async Chunking

import asyncio
from chonkie import SemanticChunker

async def process():
    chunker = SemanticChunker(chunk_size=512)
    chunks = await chunker.achunk("Document text...")
    return chunks

asyncio.run(process())

Advanced Topics

All Chunkers: Detailed guide to all 12 chunking strategies including Token, Fast, Sentence, Recursive, Semantic, Late, Code, Neural, Slumber, Table, and TeraflopAI → Chunkers

Refineries and Pipeline API: OverlapRefinery, EmbeddingsRefinery, CHOMP architecture, pipeline methods, validation rules, and best practices → Refineries and Pipeline

Integrations: Embedding providers (9+), LLM Genies (5+), Vector database handshakes (9+), Porters, Tokenizers, Chefs, Fetchers → Integrations

API Server and JavaScript SDK: Self-hosted REST API with chonkie serve, Docker deployment, pipeline persistence, and the @chonkiejs JavaScript packages → API Server and JavaScript