SuperMew
A knowledge-base-first agent platform, built around persistent threads, resumable runs and immutable document versions, that provides RAG, human-in-the-loop review and audited skill and tool execution.
SuperMew models a chat request not as a single unrecoverable HTTP call but as a set of persistent, replayable domain objects: threads hold long conversation boundaries, runs represent a resumable execution, an event journal records what happened, and checkpoints let a run pause for human review and resume later without losing identity.
Its retrieval pipeline combines dense and Milvus-native BM25 search through reciprocal rank fusion and auto-merging, with optional reranking and up to one step-back or HyDE rewrite, and documents are published through a two-phase build so an in-progress index update never affects what is currently searchable. Failures in retrieval or reranking degrade explicitly rather than silently, and are recorded in a trace. The backend is FastAPI, SQLAlchemy, PostgreSQL, Redis and Milvus, with a Vue 3 and TypeScript frontend.