Getting started¶
Refindery runs as a single local process. The only hard external dependency is an embedding provider (for indexing and search) and a vector store. You choose the vector store — and therefore how much infrastructure you run — with one setting:
-
Daemon-free (LanceDB)
Everything in one process, all state under
data/. No Docker, no daemon. The fastest way to try Refindery and the recommended workstation profile.REFINDERY_VECTOR_STORE=lancedb -
Docker (Qdrant)
Qdrant runs as a daemon (native server-side hybrid fusion and filter pushdown). The default for larger collections. Run only Qdrant in Docker, or the whole stack.
REFINDERY_VECTOR_STORE=qdrant
Both stores pass the same conformance suite, so you can start daemon-free and move to Qdrant later by registering and backfilling into the new store.
Pick your path¶
| You want to… | Start here |
|---|---|
| Try it on macOS with the least friction | Installation → macOS one-stop (no Docker) |
| Run the full Qdrant stack on macOS | Installation → macOS one-stop Docker |
| Set it up by hand on any OS | Installation → Manual minimal profile |
| See it actually work end-to-end | Quickstart |
| Confirm the install is healthy | Validate the install |
Prerequisites¶
- Python 3.13+
- uv for dependency management
- An embedding provider API key (the default is Voyage; Cohere, OpenAI, and local models are also supported). Indexing and search do not work until a provider is configured.
- Docker only if you choose the Qdrant profile.
Once installed, head to the Quickstart to ingest your first page and run a search.