What's new in AI Workbench 0.5.4

0.5.4 is a beta-feedback release on the 0.5.0 Enterprise Access Control line — every change responds to first-wave beta feedback (thanks, David Jones-Gilardi). It adds bulk document deletion and parallel ingest to the Knowledge Base Explorer, and fixes three things that broke the Docker quickstart on the first run. There is one additive wire-contract change (the new bulk-delete endpoint), no breaking change, and no data migration.

Bulk select + delete in the Knowledge Base Explorer

Clearing 100 documents used to take 100 trash clicks and 100 confirmation dialogs. The document table now has a checkbox column with select-all scoped to the visible/filtered rows — a filter narrows the blast radius, never widens it — plus a "Delete selected (N)" action bar and a single confirmation for the whole batch.

Server-side, the new POST /api/v1/workspaces/{ws}/knowledge-bases/{kb}/documents/bulk-delete endpoint takes { documentIds: string[] } (1–100 ids per call; the web app pages larger selections). Bulk is not a side door around row-level policy: each id runs the same RLAC mutation gate, the same chunk cascade, and writes the same audit record as the single DELETE. Per-id failures land in a failed array without aborting the rest of the batch, and the web app reports partial failures in a warning toast.

Parallel ingest in the queue dialog

The ingest queue drained one file at a time, even though the runtime already bounds concurrent ingest jobs server-side (default 4). The queue now keeps up to N jobs in flight — a "Parallel ingests" picker (1/2/4/8, default 4) in the queue header — and every running row streams its own live progress. Set it to 1 to restore the old sequential behavior for rate-limited embedding providers. Duplicate/name-conflict prompts still surface one at a time, and a failing file fails independently without tanking the batch.

The Docker quickstart works on the first run

Three fixes, all from the same beta pass:

• Writable data volume. The image runs as the non-root node user but never created /var/lib/workbench, so Docker initialized the named volume root-owned and the first workspace write failed with EACCES. The mount point is now pre-created with the right ownership — no user: "0:0" override, no manual chown.
• Ollama on the host is reachable. The base URL was hardcoded to localhost, which inside a container is the container itself. The runtime now honors OLLAMA_BASE_URL (compose defaults it to http://host.docker.internal:11434/v1, with a host-gateway mapping that also covers Linux Engine), the LLM service form gains an Endpoint base URL field, and transport errors now name the endpoint they tried. See "Ollama on the host" in docs/docker.md.
• Mock workspaces ingest out of the box. Mock workspaces were seeded with a credential-less NVIDIA embedding service, so the first ingest failed with 400 embedding_unavailable. They now seed the credential-free mock embedder, so the zero-credential demo flow works end to end.

See CHANGELOG.md for the full 0.5.4 entry, and docs/whats-new-0.5.0.md for the narrative tour of the Enterprise Access Control release this line builds on.