// the take
serviette is a thin, no-code wrapper around Pathway's Live Data Framework that turns a folder of documents plus one YAML file into a RAG chat and REST API, with changes to the source documents reflected in answers within seconds. It targets anyone who wants a self-hosted Q&A server over private docs without writing retrieval code or picking a re-indexing schedule — eight vector DB backends are supported, with DuckDB embedded as the zero-setup default.
The indexer/server split is real architecture, not marketing: they're separate processes sharing only the vector DB, so the server stays stateless and horizontally scalable while the indexer shards independently, and either side can restart without taking the other down. Every one of the eight backends has an integration test running the same index/retrieve/delete scenario against a real Docker instance (not mocks), which is more backend-rigor than most RAG libraries bother with. It builds on Pathway's existing parsers/splitters/embedders instead of reimplementing them, so multimodal ingestion (PDFs, Office docs, OCR, audio/video) comes for free rather than being bolted on. The benchmark write-ups are unusually honest about their own limits — they flag proxy measurements, single-run variance, and where the FRAMES numbers are paired deltas rather than absolute scores.
It's a wrapper over a third-party engine (Pathway) that requires its own free license key just to run — one more external account and one more thing that can gate your self-hosted stack from working. The default DuckDB backend is explicitly single-writer, so the zero-setup path isn't the path you'd actually run in production with concurrent writers. No built-in auth and no conversation memory or user accounts — the README is upfront that these are missing, but that still means you're bolting on session handling and an authenticating reverse proxy yourself before this is usable by more than one person at a time. Feature parity across backends isn't uniform: Pinecone loses both hybrid search and the document-catalog questions because it can't enumerate its own vectors, so swapping backends can silently remove functionality you were relying on.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →