ADR-D15: protea-method distribution channels

Status:

Accepted

Date:

2026-05-05

Phase:

F-OPS

Context

protea-method is the pure inference path (KNN, feature compute, re-ranker apply). Because it has no FastAPI / SQLAlchemy dependencies it can ship independently of the platform, addressing audiences that want the method without operating the full stack.

Decision

Three distribution channels:

  • PyPI public package (pip install protea-method).

  • Docker Hub minimal image (protea-method-runtime, ~3-4 GB with one canonical PLM and one default booster, CLI protea-predict input.fasta output.tsv).

  • Companion engineering paper covering the F-OPS pillars (containers, multi-target deployment, observability, secrets).

Consequences

  • Two release surfaces to maintain.

  • LAFA submission containers (D23) build on top of protea-method-runtime.

  • Engineering paper aligned with F-OPS deliverables.

Resolution

Closed.