finds.dev

isaacus-dev/semchunk

A fast, lightweight and easy-to-use Python library for splitting text into semantically meaningful chunks.

View on GitHub

★ 669PythonMITupdated Jun 2026

// the take

semchunk is a Python text-chunking library that splits documents into token-bounded pieces for RAG pipelines, using a recursive separator-based algorithm by default and an optional paid Isaacus API for AI-driven chunking. It's aimed at anyone building retrieval pipelines who wants chunk sizes measured in actual tokens rather than characters, with support for any tokenizer (Tiktoken, Transformers, or a custom callable).

Tokenizer-agnostic by design — you can plug in Tiktoken, Transformers, or any custom token-counting function, so it fits into an existing pipeline instead of forcing one. Practical production features are built in rather than bolted on: offset tracking, overlap control, multiprocessing for batch jobs, and a memoized token counter with a configurable cache size. It's a genuinely small, auditable codebase (one core module) with CI and Codecov wired up, and it's already load-bearing in other real projects (Docling, Microsoft's Intelligence Toolkit), which is a better maturity signal than stars alone.

The headline '15% better RAG performance' claim comes entirely from the maintainer's own blog post and a single benchmark dataset (Legal RAG QA) — treat it as marketing until someone outside Isaacus reproduces it. The 'AI-powered chunking' mode that the README leads with isn't actually part of this library's free functionality — it requires an Isaacus API key and calls their paid enrichment model, so the default open-source behavior is really just recursive separator splitting, comparable to what LangChain's splitter already does. It's a single-maintainer project (Umar Butler), which is a real bus-factor risk if you're depending on it for a production RAG pipeline. The recursive splitting heuristics (sentence terminators, clause separators, etc.) are tuned for general prose and legal text — there's no stated handling for structured content like code, tables, or markdown headers, so chunk quality on non-prose documents is unproven.

We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →

// at a glance

Stars
669
Forks
48
Language
Python
License
MIT
Last commit
13 Jun 2026

// nearby

If semchunk isn't quite it.

All repos →