zeroentropy-ai/zchunk

Leverages Llama-70B's logprobs to detect semantic boundaries without regex or embedding similarity—by injecting a special token and measuring model confidence in its placement across the document. Includes a benchmarked hyperparameter tuning pipeline and evaluates performance on LegalBenchConsumerContractsQA using dual metrics: retrieval faithfulness and signal-to-noise ratio. Outperforms fixed-size and embedding-based chunking strategies across multiple document types (Python, Markdown, HTML, legal text) with optimization for KV-cache inference efficiency.

254 stars. No commits in the last 6 months.