skygazer42/MimirQ

What it solves

MimirQ is designed for enterprise-grade RAG (Retrieval-Augmented Generation) where the primary challenge is not just vectorizing documents, but ensuring that errors are locatable, strategies are replaceable, and quality is regression-tested. It addresses the "black box" problem of simple RAG pipelines by providing a controllable knowledge pipeline where every step—from parsing and cleaning to chunking and retrieval—can be inspected, audited, and versioned.

How it works

The system implements a structured knowledge pipeline: Data Evaluation $\rightarrow$ Scenario-based Parsing $\rightarrow$ Cleaning/Governance $\rightarrow$ Business Chunking $\rightarrow$ Indexing $\rightarrow$ Hybrid Retrieval $\rightarrow$ Reranking $\rightarrow$ Golden Regression.

It supports a vast array of modular components that can be swapped based on the document type or business need:

• Parsing: Over 30 backends including DeepDoc, Docling, MinerU, and Marker.
• Chunking: 86 different strategies (semantic, recursive, parent-child, RAPTOR, etc.) with visual previews.
• Retrieval: Hybrid search combining Milvus, FAISS, or Chroma with BM25, SPLADE, and ColBERT.
• Evaluation: A "Golden Dataset" approach for regression testing, using metrics like Recall and MRR to ensure quality doesn't drop between versions.
• Integration: It can act as a standalone system or as a specialized RAG layer for Dify via External Knowledge APIs.

Who it’s for

It is built for teams delivering professional knowledge bases (such as government services) who need high precision, strict audit trails, and the ability to perform rigorous quality assurance before deployment.

Highlights

• Extreme Modularity: Swap between dozens of parsers, chunking strategies, and rerankers without rebuilding the entire pipeline.
• Observability: Visual tools to inspect chunks, retrieval traces, and reranking processes.
• Enterprise Governance: Includes document-level ACL (Access Control Lists), security trimming, and SAML SSO.
• Regression Testing: Built-in "Golden" question sets to prevent quality regressions during updates.
• Dify Compatibility: Native support for Dify's External Knowledge API, allowing it to serve as the high-precision retrieval engine for Dify workflows.