We’re pleased to announce the release of Recotem 2.0.0.
This is a major release that completely re-architects Recotem. The previous multi-service web application (Django / DRF / Vue / Celery, with a database and message broker) has been replaced by a single, recipe-driven Python package and Docker image. A single YAML recipe now defines where your data lives, how the model is trained, and where the result is written — no database, broker, or admin UI required.
Highlights
- Recipe-driven workflow — 1 YAML recipe = 1 model = 1 API endpoint. The recipe is the single source of truth for data, training, and output.
- Two simple commands — recotem train <recipe.yaml> produces a signed model artifact, and recotem serve --recipes exposes it as a FastAPI service under /v1/recipes/{name}:recommend (plus :recommend-related and :recommend-batch).
- Pluggable data sources — built-in CSV, Parquet, BigQuery, and SQL (PostgreSQL / MySQL / SQLite) sources, plus a plugin contract for your own.
- Automated tuning — Optuna-driven hyperparameter search across irspack (GitHub - tohtsky/irspack: Train, evaluate, and optimize implicit feedback-based recommender systems. · GitHub) algorithms, with optional per-algorithm trial budgets.
- Signed, hot-swappable artifacts — HMAC-signed binary artifacts with multi-key rotation; serve watches the recipe directory and hot-swaps models when a new artifact appears, with no downtime.
- Security-hardened by default — SSRF-guarded fetching with mandatory SHA-256 pinning, an explicit path-scheme allow-list, restricted env-var expansion, and
structured-log redaction of keys and credentials. - Production-ready deployment — multi-stage Docker image, Docker Compose tutorial, Helm chart with an optional training CronJob, Kubernetes manifests, and an optional
Prometheus /metrics endpoint.
Upgrading from 1.x
Recotem 2.0.0 is a ground-up rewrite and is not backward compatible stack. There is no in-place upgrade path; please re-create yourworkflow as a recipe. See the Getting Started guide below to get going quickly.
Resources
- Document: https://recotem.org/
- GitHub release: Release v2.0.0 · codelibs/recotem · GitHub
- Repository: GitHub - codelibs/recotem: An easy to use interface to recommender systems. · GitHub
- PyPI: Client Challenge
- Docker image: Package recotem · GitHub
Install
pip install recotem
# or
docker pull ghcr.io/codelibs/recotem
Thank you for using Recotem!