Show HN: Built a tool solve the nightmare of chunking tables in PDF vs. Markdown (github.com)

A developer announced an open-source, production-ready ingestion toolkit that fixes a common RAG pain point: static, one-size-fits-all chunking. Instead of blindly slicing documents into fixed blocks, the tool uses layout-aware parsing (Docling) and file-type-aware heuristics to chunk text differently for PDFs, code, research papers, and Markdown. Crucially, it preserves table structure by converting PDF tables to Markdown before chunking, maintaining relational data that often gets lost with naive splitting. The project is lightweight, extracted from a battle-tested private RAG platform, designed to run on your own hardware for privacy-conscious deployments. Why this matters: smarter chunking yields better context windows for embeddings and retrieval, reducing degraded answers and hallucinations that arise from poor document segmentation—especially with tables, structured content, and source code. Technical highlights include layout-aware parsing, separate chunking strategies per file type, table-to-Markdown preservation, minimal dependencies, and an upcoming pip package. The repo is open for issues and contributions, making it a practical option for teams wanting improved, privacy-first RAG ingestion without heavy infrastructure changes.