Records

A record carries the text and vector field values for one key, plus optional metadata. Build one with the chained Record builder, then stage it with a Writer.

import numpy as np
from valise import Record

rec = (
    Record()
    .text("body", "the quick brown fox")
    .vector("dense", np.zeros(64, dtype=np.float32))
    .at(1_700_000_000)        # created_at, unix seconds
    .child_of("parent-key")   # link as a child of another record
)

with store.writer() as w:
    w.put("kb", "a", rec)
    w.commit()

• text(field, value) — set a text field.
• vector(field, values) — set a vector field; values is a 1-D float32 ndarray, borrowed zero-copy by the native layer.
• at(unix_secs) — set the record timestamp. The engine requires created_at to be non-decreasing in commit order; out-of-order timestamps in one commit error at commit time.
• child_of(parent) — link this record under a parent key (str | int | bytes).

Keys

A key is str, int, or bytes. Booleans and negative integers are rejected.

Stored records

Reader.get / Reader.get_many (and Store.get) return a Stored dataclass:

s = store.reader().get("kb", "a")
s.key          # the record key
s.collection   # owning collection
s.created_at   # unix seconds
s.text         # the text field value, or None
s.vectors      # list[(field_name, np.ndarray)] — the moved arrays, not copied

Lossy vector codec

Vectors are stored through a quantization codec (QAM or UPQ, chosen per field in the schema), which is lossy. A round-tripped vector preserves its dimension and finiteness, but not its exact values. Do not assert that a fetched vector equals the one you stored.