Vector

Use Vector for fixed-width numeric vectors supplied by another system: precomputed embeddings, dense descriptors, sensor windows, or model features.

{
  "embedding": [0.12, -0.03, 0.44, 0.08]
}

import json2vec as j2v

embedding = j2v.Vector(
    "embedding",
    n_dim=4,
    objective="l2",
)

Use Vector when embeddings or dense features are already present in the input record. Use Text when json2vec should compute embeddings from strings. Use an Array when repeated measurements have item structure.

Input Values

Vector accepts:

• Python lists or tuples.
• 1D NumPy arrays.
• 1D PyTorch tensors.

Every valued input must have exactly n_dim numeric elements. None is encoded as a null state, and missing array positions are encoded as padded state. Prepare vector scale and normalization upstream unless a custom tensorfield owns that behavior.

Examples

Common vector fields include:

• Precomputed text, image, audio, or product embeddings.
• Dense features emitted by a fraud, ranking, search, or recommendation model.
• Sensor windows or compact numerical descriptors.
• External representation vectors that should be joined with structured fields.

Configuration

Option	Default	Notes
n_dim	required	Vector width. Must be positive.
objective	"l2"	Content reconstruction objective: "l1" or "l2".

Target Behavior

When a Vector is masked or used as a target, the decoder predicts:

• state: probabilities for valued, null, padded, and masked.
• content: the reconstructed vector.

For non-valued predicted states, public output uses a zero vector for content so API consumers get a stable shape without treating zeros as observed values.

Tracked metrics include vector reconstruction loss, mean absolute error, and root mean squared error.

Prediction Output

Model.predict(...) returns state probabilities and reconstructed vector content:

{
    "record/embedding": {
        "state": {"valued": ..., "null": ..., "padded": ..., "masked": ...},
        "content": ...,
    }
}

Notes

Use Text when json2vec should compute text embeddings from strings. Use Vector when embeddings or dense features are already present in the input record. Specialized vector or media handling can be implemented with custom data types; the built-in package does not include image, video, or audio tensorfields.