Ali CanAli Can
GündüzGündüz

Open-source RAG toolkit: structure-aware chunking along Markdown headings, hybrid dense + BM25 retrieval fused with RRF, and a fail-open reranker.

1. chunkSplit along Markdown headings; each ~1500-char chunk keeps its heading path
2. embedTurn every chunk into a dense vector
3. retrieveDense + BM25 search, fused with reciprocal rank fusion
4. rerankRe-score the fused hits. Fails open: if it is down, results pass straight through
5. generateAnswer from the top chunks, with sources

A planned fine-tune on rented GPUs, to be published as an open release.