Vector

A fixed-width numeric vector supplied by another system.

Your application may already have a merchant representation from another system. Vector lets the payment model use that supplied numeric evidence alongside the amount and merchant name. The coordinates must have a fixed width and be available at the payment’s decision time.

import relflow as rf

model = rf.Model.xs(
    amount=rf.Number,
    merchant=rf.Category,
    merchant_embedding=rf.Vector(n_dim=4),
    disputed=rf.Boolean(mask=True),
)

One labeled payment:

amount: 84.50
merchant: North Market
merchant_embedding: [0.12, -0.03, 0.44, 0.08]
disputed: false

The four coordinates illustrate the supplied shape; set n_dim to the actual width your application supplies. Use Text to encode strings and Branch for repeated transaction records.

Input and options

Each valued input must be an Arrow list or fixed-size list with exactly n_dim numeric, non-null elements. Values convert to float32. A null vector is a separate field state; null elements inside a valued vector are invalid. A vector with a NaN or infinite element is not a measurement: it takes the other value state with zero content, for inputs and targets, so one bad element cannot poison training. relflow does not fit a vector normalizer: prepare coordinate scales upstream.

Option	Default	Meaning
n_dim	required	Positive vector width.
jitter	rf.Jitter()	Training noise in supplied coordinate units.
objective	"l2"	Reconstruction loss: "l1" or "l2".

Jitter changes finite coordinates of valued training inputs before projection. Targets remain pristine. Both settings of jitter.normalize use the same coordinate space; see Jitter.

Learning and output

Reconstruction combines state cross entropy with content loss on valued targets. MAE and RMSE metrics use the supplied coordinate scale.

In prediction output, null ancestor records preserve their structural holes. rf.Writer rejects these fixed-width vectors when PyArrow cannot round-trip them through Parquet; use Arrow IPC for affected tables.

Prediction content is fixed_size_list<float32>[n_dim]. When the most likely predicted state is non-valued, the writer emits a zero vector. Zeros therefore do not establish whether a vector was observed; read the shared state and inferred members in the prediction envelope.

Use shared leaf options to make the vector a target or export a contextual embedding. That exported embedding has the model width and is distinct from the supplied vector.