Vector

Backed by: @pithy-sh/vector, its GET {basePath}/indexes, answered by your own Worker. Scope: vector:indexes:read.

What it shows

Vector is the capability it reads, the index read is the response behind it, metadata indexes is what drifted means, and reprocessing is how an index becomes current again.

Every index your config declares, how much is in it, and whether it is current, read live from your Worker. It appears in the rail under Operations, beside Migrations and Secrets, when your Worker composes vector, at whatever path you mounted it. An index is the environment as a running system, not your customers’ business.

The index lists each index by its config name, in config order, none omitted. Each row carries its vector count, or Not described when Vectorize could not say, and one chip, the first of these that is true of it:

ChipWhen
DriftedThe provision record and config disagree on the metadata indexes
Not provisionedConfig declares the index and the provision record has none
UnreadableVectorize did not describe the index
Dimensions differThe index was created with other dimensions than config declares
Not currentDocuments are under another model, or were never embedded
CurrentEvery document is embedded under the declared model

A drifted index leads because it is the one state in which every other vector route on your Worker refuses.

The record

Declared: the model, dimensions and metric, as pithy.config.ts sets them.

Vectorize: the vector count, the dimensions the index was created with, and Writes reflected up to a time, or None processed yet when Vectorize has processed nothing. When the dimensions differ, both numbers are shown with your Worker’s own sentence.

Documents: how many are embedded under the declared model, how many under another model, and how many were never embedded, with the namespace count beside them.

Provision: provisioned, not provisioned, or drifted. A drifted index lists every missing and mis-typed metadata field. It leads the record whenever it is not provisioned.

Current is read from the counts

An index is current when nothing is under another model and nothing is unembedded. Otherwise the record names both counts and pithy vector reprocess, and when documents sit under another model it says what that costs: a search spans two embedding spaces, and nothing errors.

There is no re-embed date, because your Worker does not report one.

Unreadable is never zero

An index Vectorize could not describe shows its reason and no count: its binding is not on the Worker, its binding cannot describe an index, or the description failed. An empty count would read as an empty index. The documents by model and the namespace count come from your corpus, not from Vectorize, so they still render.

Your Worker’s sentence, not ours

A provision or dimension problem is drawn with the message your Worker sent, verbatim. Your Worker knows its environment and its config, and we know neither, so the pane writes no remedy of its own. The one command it names is pithy vector reprocess, for counts your Worker reports without a sentence.

Namespaces are a count

Never a list. A namespace is often a tenant or organization id, and your Worker sends only the count, so there is nothing here that could name one.

Read once, held nowhere

The pane reads when you open it, on the current connection, and keeps nothing. The answer is parsed with the capability’s own schema, not cached, not even for the screen’s minute, and held in no Pithy table. A response that does not parse fails on screen, naming the environment.

A connection without vector:indexes:read is refused by name, and nothing is sent. The scope is a read, so it is in the default grant; Scopes has how to add one to a connection made before it existed.

It reads, and it never writes

Provisioning an index and re-embedding a corpus are not buttons here. Both are writes against production, and both belong to the CLI: pithy provision and pithy vector.

ESC