AI Feature Discovery

Use thyme discover to automatically generate feature definitions from your data sources.

thyme discover is an AI-powered command that introspects a data source's schema and generates complete Thyme SDK code (@dataset, @pipeline, @featureset) tailored to your use case.

How it works

Schema introspection - The CLI connects to your data source and samples rows to infer the schema (column names, types, sample values).
LLM code generation - An Anthropic Claude model generates Thyme SDK code based on the schema and your stated use case.
Validation - The generated code is validated by importing it and checking for registry errors.
Output - Code is printed to stdout or written to a file.

Example: Fraud detection from Postgres

thyme discover \
    --source-type postgres \
    --use-case "fraud detection" \
    --pg-host db.your-company.com --pg-port 5432 \
    --pg-database analytics --pg-table orders \
    --pg-user readonly --pg-password $PG_PASSWORD \
    --windows "1h,24h,7d" \
    --output fraud_features.py

This produces a complete Python file with:

A @dataset for the orders table with the correct schema
A @source with PostgresSource configured
A @pipeline with aggregations relevant to fraud detection across the specified windows
A @featureset with extractors

Example: Anomaly detection from Kafka

thyme discover \
    --source-type kafka \
    --use-case "price anomaly detection" \
    --kafka-brokers kafka:9092 \
    --kafka-topic product-prices \
    --entity-key product_id \
    --windows "7d,30d,180d" \
    --output price_features.py

Supported source types

Source type	Required options
`iceberg`	`--catalog`, `--database`, `--table`
`postgres`	`--pg-host`, `--pg-database`, `--pg-table`, `--pg-user`, `--pg-password`
`s3json`	`--s3-bucket`
`kafka`	`--kafka-brokers`, `--kafka-topic`
`jsonl`	`--path`

Tips

Provide an --entity-key hint when the entity key isn't obvious from the schema.
Use --sample-n to control how many rows are sampled for schema inference (default: 5).
Review before committing - always review generated code before thyme commit.
Use --auto-commit with --output to generate and deploy in one step (for quick prototyping only).
Set ANTHROPIC_API_KEY in your environment or pass --api-key.

See the CLI Reference for the full list of options.