knowledge-rag ships local MCP RAG

The strongest part of this launch is not “local” by itself, it’s the retrieval stack discipline: hybrid search plus reranking is exactly where most RAG systems lose precision. If the ONNX/FastEmbed story holds up in practice, this could be a genuinely useful default for private knowledge workflows.

• –Cross-encoder reranking is the real differentiator here; it should clean up the noisy top-k results that sink plain vector search.
• –Hybrid BM25 + semantic + RRF is a sensible fit for technical corpora, where exact terms, acronyms, and code identifiers matter as much as embeddings.
• –Built-in retrieval evaluation is a smart touch; MRR@5 and Recall@5 make it easier to tell whether the system is actually improving instead of just feeling smarter.
• –The MCP toolset makes this more than a search library. It starts to look like a persistent knowledge layer for Claude Code, which is a more compelling use case than “yet another local RAG demo.”
• –The one caveat: the public repo surface appears to still document an older Ollama-based path, so this launch reads like a fast-moving product in transition rather than a fully settled stack.