recotem.org

What It Is

Recotem is a recommendation system tool aimed at non–machine learning specialists. Its core idea is “recipe-driven”: a YAML file defines data sources, field mappings, cleaning rules, training configuration, and model artifact output. Each recipe training run produces a signed artifact and maps to a /predict/{name} HTTP recommendation endpoint.

Core Capabilities

Based on the main content, Recotem is not focused on providing a low-level algorithm library, but rather on productizing the engineering workflow around recommendation systems. Its CLI can handle data fetching, Optuna hyperparameter tuning, training, signing, and service startup. Data sources include CSV, Parquet, BigQuery, SQL, and GA4, with plugin-based extension also supported. The server side is built on FastAPI. Training and serving processes are decoupled, with models passed through binary artifacts, so there is no dependency on a shared database or RPC.

Security and Deployment

Its security design stands out. Artifacts are signed with HMAC and fully verified before the service loads them. It supports KeyRing-based multi-key rotation, and deserialization includes an FQCN allow-list. Verification cannot be bypassed in production, and unsigned loading in development mode is also explicitly protected. For deployment, it provides Docker, Kubernetes, and cron/systemd modes, and supports hot-swapping artifacts: after a new model passes verification and deserialization, it is switched in atomically; if it fails, the old model continues serving, with the failure recorded in health checks and Prometheus metrics.

Pricing and Open Source

The main content only shows “View on GitHub” and does not clearly state the license, open-source terms, commercial edition, or hosted service pricing. As a result, its open-source boundaries and pricing model cannot be determined. For production use, you should first verify the repository license, maintenance activity, and security update process.

Pros, Cons, and Who It’s For

Its strengths are configuration-driven usage, a low barrier to entry, and clear engineering boundaries. It is well suited to teams that need to quickly turn user-item interaction data into a recommendation API. It also fits self-hosted scenarios involving periodic offline retraining and gradual model replacement. Its limitations are that the main content does not demonstrate advanced feature engineering, online learning, experimentation platform capabilities, or visual operations tools. Teams will still need some DevOps and data source configuration expertise.

Access from China

No information was found about access from mainland China, mirrors, payments, or localization support, so china_access can only be rated as unknown. If dependencies such as GitHub, BigQuery, or GA4 are difficult to access, domestic deployments may require a network proxy or replacement data sources. Alternatives to compare include LensKit, RecBole, LightFM, implicit, and Metarank.