Handwriting a RAG engine from scratch: Implementing semantic chunking and hybrid retrieval

On 2026-11-05, a Juejin article titled "Writing a RAG Engine from Scratch: Semantic Chunking and Hybrid Retrieval" walked readers through building a complete RAG engine by hand. The piece is explicit that this goes beyond simple vector search: the goal is a full pipeline with several distinct pieces. Chunking is semantic and structure-aware, operating at the Markdown hierarchy level so that documents are split along their heading and nesting structure rather than by a fixed character count. Keyword search is implemented by writing the BM25 algorithm manually instead of calling a library. Hybrid retrieval is built by combining the BM25 rankings with dense vector search results through reciprocal rank fusion (RRF), which merges two ranked lists without needing normalised scores. For multi-turn question answering, the article adds a rewriting step that uses an LLM to rewrite follow-up questions into standalone ones, so a query like "what about the second part?

" can be resolved on its own. Finally, a quality evaluation system is included to score retrieved results. As of the only available report, the article remains at the feature-overview stage. It does not provide the chunking heuristics, BM25 source code, RRF parameters, retrieval or end-to-end benchmarks, evaluation metrics, or any follow-up updates. Readers interested in the actual implementation details have to wait for more material.