Recotem

English

日本語

English

日本語

Appearance

Overview ​

Recotem trains and serves a recommender model from a single small YAML configuration file. No machine-learning expertise required.

You describe your data source, a few training preferences, and where to save the result. Recotem handles the rest: fetching the data, finding the best algorithm, training the model, and answering recommendation requests over HTTP.

We call that configuration file a recipe — it describes what data to use, how to train, and where to serve. One recipe produces one model and one API endpoint.

No database, no message broker, no administration interface. A recipe file, two commands, and an HTTP endpoint.

How it works ​

Recotem is built around two processes that communicate only through a signed binary file called an artifact (the trained model):

  1. recotem train reads your recipe, fetches the interaction data (purchases, clicks, reads), searches for the best algorithm and settings, trains the final model, and writes a signed artifact to disk or cloud storage.

  2. recotem serve watches the artifact directory and loads each model as a REST endpoint at /predict/{name}. When recotem train produces a new artifact, the server picks it up automatically — no restart needed.

Because the two processes share nothing except the artifact file, they can run on different machines. A nightly batch job can write to an S3 bucket while a long-running server reads from the same bucket, hot-swapping the model as soon as a fresh one appears.

recipe.yaml  →  recotem train  →  artifact.recotem  →  recotem serve
                (batch job)        (HMAC-signed)        (FastAPI, hot-swap)

any scheduler      local FS / S3 / GCS          POST /predict/{name}

The artifact is protected by an HMAC signature (a tamper-detection code). The serving process verifies the signature before loading any model, so a corrupt or altered file is rejected rather than used.

Who Recotem is for ​

Recotem is built for teams that want to add recommendations to their product without a dedicated machine-learning team or a complex data infrastructure. If you can export interaction data to a CSV or a BigQuery table, Recotem can turn it into a working recommendation API.

It is a good fit when:

  • Your team does not include ML engineers but you still need recommendations in your app or service.
  • You want a working recommendation endpoint running in under an hour, not weeks.
  • You want to retrain on a schedule (nightly, weekly) with minimal operational overhead.
  • Your data lives in CSV files, Parquet files, or BigQuery — and you want to serve predictions over HTTP without managing a separate database or message queue.

The three-step mental model ​

  1. Write a recipe. A small YAML file describes where your data lives, which columns identify users and items, and where to save the trained model.

  2. Run recotem train. One command fetches data, searches for the best algorithm and settings, trains the final model, and writes a signed artifact.

  3. Run recotem serve. One command starts an HTTP server that loads the artifact and answers POST /predict/{name} requests. Retrain and the server updates itself.

Next steps ​