chunklet 1.0.4.post4

A smart multilingual text chunker for LLMs, RAG, and beyond.

📦 Chunklet: Smart Multilingual Text Chunker

Chunk smarter, not harder — built for LLMs, RAG pipelines, and beyond.
Author: Speedyk_005
Version: 1.0.4.post4 (🎉 first stable release)
License: MIT

🚀 What’s New in v1.0.4.post4 (with cli support)

• ✅ Stable Release: v1.0.4 marks the first fully stable version after extensive refactoring.
• 🔄Multiple Refactor Steps: Core code reorganized for clarity, maintainability, and performance.
• ➿ True Clause-Level Overlap: Overlap now occurs on natural clause boundaries (commas, semicolons, etc.) instead of just sentences, preserving semantic flow better.
• 🛠️ Improved Chunking Logic: Enhanced fallback splitters and overlap calculations to handle edge cases gracefully.
• ⚡ Optimized Batch Processing: Parallel chunking now consistently respects token counters and offsets.
• 🧪 Expanded Test Suite: Comprehensive tests added for multilingual support, caching, and chunk correctness.
• 🧹 Cleaner Output: Logging filters and redundant docstrings removed to reduce noise during runs.

---

🔥 Why Chunklet?

Feature	Why it’s elite
⛓️ Hybrid Mode	Combines token + sentence limits with guaranteed overlap — rare even in commercial stacks.
🌐 Multilingual Fallbacks	CRF > Moses > Regex, with dynamic confidence detection.
➿ Clause-Level Overlap	overlap_percent now operates at the clause level, preserving semantic flow across chunks using , ; … logic.
⚡ Parallel Batch Processing	Multi-core acceleration with mpire.
♻️ LRU Caching	Smart memoization via functools.lru_cache.
🪄 Pluggable Token Counters	Swap in GPT-2, BPE, or your own tokenizer.

---

🧩 Chunking Modes

Pick your flavor:

• "sentence" — chunk by sentence count only
• "token" — chunk by token count only
• "hybrid" — sentence + token thresholds respected with guaranteed overlap

---

📦 Installation

Install chunklet easily from PyPI:

pip install chunklet

To install from source for development:

git clone https://github.com/speed40/chunklet.git
cd chunklet
pip install -e .

---

💡 Example: Hybrid Mode

from chunklet import Chunklet

def word_token_counter(text: str) -> int:
    return len(text.split())

chunker = Chunklet(verbose=True, use_cache=True, token_counter=word_token_counter)

sample = """
This is a long document about AI. It discusses neural networks and deep learning.
The future is exciting. Ethics must be considered. Let’s build wisely.
"""

chunks = chunker.chunk(
    text=sample,
    mode="hybrid",
    max_tokens=20,
    max_sentences=5,
    overlap_percent=30
)

for i, chunk in enumerate(chunks):
    print(f"--- Chunk {i+1} ---")
    print(chunk)

---

🌀 Batch Chunking (Parallel)

texts = [
    "First document sentence. Second sentence.",
    "Another one. Slightly longer. A third one here.",
    "Final doc with multiple lines. Great for testing chunk overlap."
]

results = chunker.batch_chunk(
    texts=texts,
    mode="hybrid",
    max_tokens=15,
    max_sentences=4,
    overlap_percent=20,
    n_jobs=2
)

for i, doc_chunks in enumerate(results):
    print(f"\n## Document {i+1}")
    for j, chunk in enumerate(doc_chunks):
        print(f"Chunk {j+1}:\n{chunk}")

---

⚙️ GPT-2 Token Count Support

from transformers import GPT2TokenizerFast
tokenizer = GPT2TokenizerFast.from_pretrained("gpt2")

def gpt2_token_count(text: str) -> int:
    return len(tokenizer.encode(text))

chunker = Chunklet(token_counter=gpt2_token_count)

---

🧪 Planned Features

[x] CLI interface with --file, --mode, --overlap, etc. [ ] code splitting based on interest point [ ] PDF splitter with metadata [ ] Named chunking presets (conceptually "all", "random_gap") for downstream control

---

🌍 Language Support (30+)

• CRF-based: en, fr, de, it, ru, zh, ja, ko, pt, tr, etc.
• Heuristic-based: es, nl, da, fi, no, sv, cs, hu, el, ro, etc.
• Fallback: All other languages via smart regex

---

💡Projects that inspire me

Tool	Description
Semchunk	Semantic-aware chunking using transformer embeddings.
CintraAI Code Chunker	AST-based code chunker for intelligent code splitting.

---

🤝 Contributing

1. Fork this repo
2. Create a new feature branch
3. Code like a star
4. Submit a pull request

---

📜 License

MIT License. Use freely, modify boldly, and credit the legend (me. Just kidding!)