// the find
superduper-io/superduper
Superduper: End-to-end framework for building custom AI applications and agents.
Superduper is a Python framework that lets you attach ML models and AI pipelines directly to your existing database — MongoDB, Postgres, Snowflake, Redis — so that inserts trigger inference automatically. The pitch is 'bring your database, not a new vector DB,' which is genuinely interesting if you're already invested in one of the supported backends. It's aimed at teams who want to add RAG or embedding pipelines without standing up a separate ML serving stack.
The plugin architecture is clean: each backend and model provider is its own installable package, so you don't drag in torch and transformers just to use OpenAI. The change-data-capture listener pattern (database insert → model inference → store result) is a real architectural win for keeping ML outputs fresh without a separate orchestrator. Cohere, Jina, and sentence-transformers plugins all ship with VCR cassettes for offline testing, which shows actual discipline. Supporting Snowflake as a data backend is a legitimate differentiator — most RAG frameworks don't go near enterprise data warehouses.
The abstraction layer is thick and leaky: you write Superduper's query DSL on top of your existing DB's query language, which means debugging failures requires understanding two layers simultaneously. The README is almost entirely navigation links with no code example — for a framework this complex, that's a real onboarding problem. Redis is listed as a supported backend but the plugin has no README, no tests, and only a stub data_backend.py, so its production readiness is unclear. Distributed compute support exists but appears to be a thin local-execution shim in practice; the 'distributed-ml' topic in the repo description is doing a lot of work it hasn't quite earned yet.