A RAG pipeline that treats debugging like a feature, not an apology

What it does\nMimirQ is an open-source enterprise RAG pipeline that treats document parsing, chunking, retrieval, and reranking as independently inspectable stages rather than a black box. It wraps around 30 document parsers, 86 chunking strategies, and 13 rerankers behind a unified interface, then tracks each step with traceable inputs, outputs, and versioned quality gates. The system is designed to plug into existing orchestrators like Dify, acting as a governable knowledge layer rather than a monolithic chatbot.\n\nThe interesting bit\nMost RAG tools optimize for the demo; MimirQ optimizes for the post-mortem. It ships with a fixed 800-question "Golden" test set to regression-test releases, and lets teams swap parsers or chunkers per document type instead of applying one-size-fits-all defaults. The philosophy is almost anti-low-code: every strategic choice is exposed so you can audit exactly where an answer went wrong.\n\nKey highlights\n- Modular pipeline from parsing to citation: swap DeepDoc, Docling, MinerU, or a dozen other parsers per document type without rebuilding the stack.\n- 86 chunking strategies and 13 rerankers, selectable by material rather than global config.\n- Built-in Golden regression suite with 800 questions to gate releases by recall and MRR.\n- First-class Chinese document support, born from real government knowledge base deployments.\n- Plugs into Dify via External Knowledge API, so you keep Dify’s workflow canvas while outsourcing retrieval governance.\n\nCaveats\n- The authors explicitly warn that simple, low-code use cases are better served by Dify or FastGPT; MimirQ is overkill if you don’t need component-level audits.\n- Optional high-fidelity parsers demand serious hardware—olmOCR suggests 48 GiB of VRAM, and the base deployment starts at 16 GB RAM.\n- With 30 parsing backends and 86 chunking strategies on offer, the configuration matrix is vast; the README notes that risky materials still require human verification.\n\nVerdict\nTeams running Chinese-language enterprise knowledge bases who need to audit, swap, and regression-test every stage of the RAG pipeline should look here. If you just want a quick "upload PDF and chat" setup, the README itself will tell you to keep walking.