“One library to split them all: Sentence, Code, Docs”

Hey! Welcome. Let's make some text chunking magic happen.

-- documentation site --

Why Smart Chunking? (Or: Why Not Just Split on Character Count?)

You could split your text by character count or random line breaks. But that's like trying to cut a wedding cake with a chainsaw. 🎂

Dumb splitting causes problems:

• Mid-sentence surprises: Your thoughts get chopped mid-way, losing all meaning
• Language confusion: Non-English text and code structures get treated the same
• Lost context: Each chunk forgets what came before

Smart chunking solves this by:

• Smart limits: respects both natural boundaries (sentences, paragraphs, sections) AND configurable limits (tokens, lines, functions)
• Language-aware: detects language automatically and applies the right rules (60+ languages supported)
• Context preservation: overlap between chunks, rich metadata (source, span, document structure)

🤔 So What's Chunklet-py Anyway? (And Why Should You Care?)

Chunklet-py is a developer-friendly text splitting library for devs, researchers, and AI engineers. It goes way beyond basic character counting. I built this because I was tired of terrible chunking options. Chunklet-py intelligently chunks text, documents, and code into meaningful, context-aware pieces, perfect for RAG pipelines and LLM applications.

Key features:

• Composable constraints: mix and match limits (sentences, tokens, sections) to get exactly the chunks you need
• Pluggable architecture: swap in custom token counters or document processors
• Rich metadata: every chunk comes with source references, spans, and structural info
• Multi-format support: PDF, DOCX, EPUB, Markdown, HTML, LaTeX, ODT, CSV, Excel, and plain text

Available tools:

• SentenceSplitter: lightweight sentence tokenization
• DocumentChunker: natural language with semantic boundaries
• CodeChunker: language-aware code chunking
• SelfTuningChunker: for mixed text/code corpora
• ChunkVisualizer: interactive web-based exploration

Perfect for prepping data for LLMs, building RAG systems, or powering AI search - Chunklet-py gives you the precision and flexibility you need across tons of formats and languages.

And that's just the start - there's plenty more to explore!

📦 Installation

Ready to get Chunklet-py running? Awesome! Let's get you set up quickly and painlessly.

The Quick & Easy Way

The simplest way to get started is with pip:

# Install and check it's working
pip install chunklet-py -U
chunklet --version

That's it! You're all set to start chunking.

Extra Features (Optional)

Want to unlock more Chunklet-py superpowers? Add these optional dependencies based on what you need:

• Structured Documents: For handling .pdf, .docx, .epub, .eml, .pptx, and other document formats:

  pip install "chunklet-py[struct-doc]"
  

• Code Chunking: For Language-agnostic code chunking features:

  pip install "chunklet-py[code]"
  

• Self-Tuning Chunking: For self-tuning mixed text/code chunking (bundles struct-doc, code):

  pip install "chunklet-py[self-tuning]"
  

• Auto Language Detection: For automatic language detection (lang="auto"). Uses py3langid to detect the language of your text:

  pip install "chunklet-py[lang-detect]"
  

• Indic Languages: For splitting text in Indic languages (Assamese, Gujarati, Kannada, Nepali, Odia, Punjabi, Sanskrit, Tamil, Telugu). Note: Bengali, Malayalam, and Hindi are already covered by the default install:

  pip install "chunklet-py[indic]"
  

• Visualization: For the interactive web-based chunk visualizer:

  pip install "chunklet-py[viz]"
  

• All Extras: To install all optional dependencies:

  pip install "chunklet-py[all]"
  

The From-Source Way

Prefer building from source? You can clone and install manually for full control:

git clone https://github.com/speedyk-005/chunklet-py.git
cd chunklet-py
pip install .[all]

(But honestly, the pip way is usually way easier!)

Want to Help Make Chunklet-py Even Better?

That's awesome! We'd love to have you contribute. Check out our Contributing Guide first, then set up your development environment:

git clone https://github.com/speedyk-005/chunklet-py.git
cd chunklet-py
# For basic development (testing, linting)
pip install -e ".[dev]"
# For documentation development
pip install -e ".[docs]"
# For comprehensive development (including all optional features like document and code chunking + docs dependencies)
pip install -e ".[dev-all]"

These install Chunklet-py in "editable" mode so your code changes take effect immediately. The different options give you just the dependencies you need.

Go forth and code! (And remember, good developers write tests. We appreciate excellent code examples!)

Quick Reference 🛠️

The Constraint-Based Logic

Chunklet-py is basically a "choose your own adventure" for data. It's constraint-based, meaning you can swap, combine, or ignore the limits below as you see fit.

The Golden Rule: You must provide at least one constraint, or the chunker has no idea when to stop.

Core Imports

Pick your weapon based on whatever data mess you're currently cleaning up.

from chunklet import DocumentChunker  # For PDFs, DOCX, and general text chaos
from chunklet import CodeChunker  # For source code (it actually respects brackets)
from chunklet import SelfTuningChunker  # For mixed text + code, self-tuning chunking
from chunklet import SentenceSplitter  # For when you just need to split sentences
from chunklet import visualizer  # Web-based chunk visualizer

Configuration & Limits

These tools don't share arguments, so don't try to use max_functions on a PDF unless you want to see a very confused Python interpreter.

DocumentChunker (Text & Docs)

Perfect for natural language where you don't want to cut someone off mid-sentence.

# Constraints are set once, at construction -- then reused by every call
chunker = DocumentChunker(
    lang="en",  # Required. Pass "auto" to detect it (needs the [lang-detect] extra)
    max_sentences=3,  # Stop after X sentences
    max_tokens=500,  # Don't blow up the LLM context
    max_section_breaks=2,  # Respect the Markdown headers
    token_counter=...,  # Required whenever you set max_tokens
    overlap_percent=20,  # Give it some "memory" of the last chunk
)

chunks = chunker.chunk_text(text)

# They're plain attributes, so tweak them whenever you like
chunker.max_sentences = 5

CodeChunker (Source Code)

Logic-aware. It doesn't do "overlap" because duplicate code is a hallucination waiting to happen.

# Again, use whichever constraints make sense for your file
chunker = CodeChunker(
    max_lines=50,  # Height limit
    max_tokens=512,  # Width limit
    max_functions=1,  # One function per chunk (keeps things tidy)
    token_counter=...,  # Required whenever you set max_tokens
)

# `strict` stays per-call, since it depends on the source
chunks = chunker.chunk_text(
    text,
    strict=True,  # True: Crash on big blocks; False: Slice 'em up anyway
)

SelfTuningChunker (Mixed Text & Code)

Detects each source as document or code, learns per-profile structural metrics (KAMA), and sizes chunks to match — no constraint tuning needed.

# token_counter is optional if you don't want max_tokens
chunker = SelfTuningChunker(lang="en", token_counter=...)

# Learning happens per call -- no queue, no separate processing step
chunks = chunker.chunk_file("docs/guide.md")
chunks += chunker.chunk_text("Some raw prose, chunked just like a file.")

# Each chunk.metadata carries "inferred_type": "document" or "code"
# Already know which? Skip the guesswork with file_type="code"
chunks = chunker.chunk_text(text, file_type="code")

The Output Object

The chunkers return a list (or generator) of Chunk objects. These are Box instances, so you can use dot notation like a civilized developer.

for chunk in chunks:
    print(chunk.content)  # The actual text/code
    print(chunk.metadata)  # Chunk metadata
    print()  # Because whitespace is free

Input Methods (Chunkers Only)

These helper methods are for the DocumentChunker, CodeChunker and SelfTuningChunker. The SentenceSplitter is a simple soul and only takes strings.

The batch methods also have show_progress, on_errors, n_jobs params.

Specialized Tools

SentenceSplitter

The "lite" version for when you just need sentences and no fancy metadata.

# Pass lang="auto" to detect it (needs the [lang-detect] extra)
# 'auto' usually guesses right, but you can specify 'en', 'es', etc.
splitter = SentenceSplitter(lang="auto")

sentences = splitter.split_text(text)

CLI (Command Line Interface)

If you prefer the terminal to an IDE, the CLI is packed with features. Just ask for help.

chunklet --help
chunklet split --help
chunklet chunk --help
chunklet visualize --help
chunklet [COMMAND] [OPTIONS*]

How Chunklet-py Compares

While there are other chunking libraries available, Chunklet-py offers versatility, performance, and ease of use in one lightweight package. Here's a quick look at how it compares to some of the alternatives:

Chunklet-py is a specialized, drop-in replacement for the chunking step in any RAG pipeline. It handles text, documents, and code without heavy dependencies, while keeping your project lightweight.

🙌 Contributors & Thanks

A huge thank you to the awesome people who helped shape Chunklet-py:

• @jmbernabotto: helped mostly on the CLI part, suggesting fixes, features, and design improvements.
• @arnoldfranz: reported the CLI Path Validation Bug (#6) that helped improve error handling.
• @AkshatSharma25: fixed EML metadata extraction to capture all header fields.

📜 License

Check out the LICENSE file for all the details.

MIT License. Use freely, modify boldly, and credit appropriately! (We're not that legendary... yet 😉)