Vector databases

Building and operating an app's vector databases from the command line, where a corpus is loaded, shaped, measured, and tuned. add loads documents, and the platform extracts, chunks, and embeds them; create, list, status, count, rm, and search cover the rest of everyday use. Configuration lives on the corpus, not in code: config shows and changes how documents are chunked and searched. Rebuilds never touch what is live: revectorize builds a new version beside the current one, search --candidate compares the two, and promote cuts over with no downtime. move places a corpus on shared or dedicated capacity.

Beyond a hand-loaded corpus, this surface runs ingestion at scale. connect points a corpus at an S3 bucket and sync pulls it in. A bulk load runs as a job: it estimates its own cost, pauses for approval over a budget ceiling, and can be paused, resumed, retried, or cancelled, with quarantine and replay for the objects it could not process. When raw objects are not usable documents on their own, a mapper shapes them: inspect reads a source's objects, map test and map deploy build and ship the transform, and remap reruns it. eval measures retrieval quality: it runs query sets against a corpus and compares runs, so you can see the numbers behind a config change.