// the take
Chonkie is a text-chunking library for Python RAG pipelines, covering everything from basic token/sentence splitting to embedding-based semantic chunking and LLM-driven agentic chunking, wrapped in a pipeline API and an optional REST server. It's for developers building document ingestion for retrieval who want more than one naive splitter without pulling in a framework like LangChain just to chunk text.
The install story is genuinely modular: the base package is a 505KB wheel that gets you token, sentence, and recursive chunking, and you only pull in heavier optional dependencies (sentence-transformers, torch-based neural models) when you actually need semantic or neural chunking. The chunker lineup spans a real cost/quality curve rather than offering one algorithm with knobs: a SIMD byte-level chunker at the fast end, embedding-based semantic chunking in the middle, and an LLM-driven agentic chunker (SlumberChunker) at the high-quality/high-cost end. The pipeline API supports both sync and async execution, so chunk → refine → embed steps can be wired into an actual ingestion service instead of glued together by hand. Built-in vector DB handshakes (pgvector, qdrant, weaviate, pinecone, etc.) remove the usual boilerplate of manually batching upserts after chunking.
The chunkers marketed as higher quality (semantic, neural, slumber) call out to an embedding model or an LLM per document, which changes the cost and latency profile completely from the token/sentence chunkers — it's easy to reach for the 'better' chunker and not notice you've added a network call and a bill per ingestion run. The benchmarks published are all about package size and raw chunking speed; there's no retrieval-quality comparison (recall/precision downstream) showing that semantic or neural chunking actually produces better RAG results than naive splitting for a given corpus, so you're picking a chunker on vibes. The new REST API server stores pipelines in a local SQLite file with no mention of auth or multi-tenancy, which is fine for a single dev box but not something you'd point multiple teams or a production ingestion job at as-is. With 45+ integrations across embedding providers, vector DBs, and LLM genies, the long tail (turbopuffer, milvus, less common embedding providers) is the kind of surface that tends to drift out of sync as upstream SDKs release breaking changes, even if the core chunkers stay stable.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →