Dataset workbench

Fine-tuning run dashboard with configuration provenance, live loss curves, validation metrics, checkpoints, cost, events, and deployment gates.

RAG knowledge-base administration with source health, sync status, chunking and embedding settings, access scope, and index progress.

Retrieval debugger that exposes ranked chunks, relevance scores, filters, answer grounding, citations, and the effective context window for a test query.

Unified image, video, and audio generation workspace with prompt direction, references, variation controls, an editable canvas, and a generation queue.

Side-by-side multimodel playground with reusable prompt messages, model and parameter controls, streaming responses, token and latency stats, and save-to-eval actions.

Prompt operations screen with structured messages, variables, version history, diffs, evaluation scores, environment promotion, and rollback context.

Inspectable response run with streaming, tool-use, waiting, completed, and failed states plus a concise reasoning summary that keeps private chain-of-thought hidden.

Tool-call and MCP control plane with server health, discoverable tools, typed request and response payloads, permission scopes, and human approval gates.

A human approval queue where each card names the automation, the action awaiting sign off, context chips, the requester, and approve or decline actions.

Vertical trigger, filter, and action step editor with dashed connectors, a Draft badge, and Save automation for a single linear flow.

Failed automation card with the error in plain words, a retry meter showing attempts used and remaining, and a payload replay affordance.

Template gallery with six categorised automation cards, use counts, a search field, and a start-from-scratch tile.