SuperMew

Why we included this project

RAG pipelines that work in a demo tend to fall apart in production: documents get updated, runs get interrupted, and you can't always say what the model actually retrieved. SuperMew treats a chat request as a set of durable domain objects rather than a throwaway HTTP call. Threads keep long conversations, runs can pause and resume mid-execution, and document versions are immutable, so rebuilding the index never pollutes what's currently searchable. Retrieval pairs dense vectors with Milvus-native BM25, then applies RRF, auto-merging, optional reranking, and evidence judging before generating an answer. It also ships a versioned evaluation harness with baselines and CI gates, plus a permission-aware tool registry covering SQL, web research, and sandboxed HTTP calls. For teams that need auditable, resumable agentic RAG rather than a quick prototype, this is a solid reference.