Skip to main content

Playground

VM-X ships with an in-app Playground for trying out completions without leaving the console. It lives at:

/workspaces/<workspaceId>/<environmentId>/playground

…and is reachable from the sidebar's "Workspaces" submenu, or from any AI Resource edit page via the Open playground button (which deep-links with the resource preselected).

Two modes​

A toggle at the top of the page picks the source of the model spec:

  • Use AI Resource — pick from the workspace's existing resources. Empty workspaces show a "Create AI Resource" CTA. Otherwise the picker pre-selects the first resource alphabetically (or the one passed via ?resourceId=<id>).

  • Use Connection / Model — type a connection name and a model name directly. The gateway parses <connection_name>/<model_name> and builds an ephemeral resource on the fly, so you can prototype against any connection without first wrapping it in a Resource.

Endpoint mode + Streaming​

A toggle group at the top of the page picks one of three input formats. Each completion can be sent through:

  • OpenAI — Chat Completions (default) — OpenAI's classic shape; supports streaming on/off, all OpenAI-compatible providers, plus multimodal attachments (image / audio / file parts).
  • OpenAI — Responses — OpenAI's newer typed-events API. The Streaming toggle is disabled in this mode (single-shot non-streaming).
  • Anthropic — Messages — drop-in compatibility with Anthropic's POST /v1/messages. Same streaming caveat as Responses (toggle disabled).

All three go through the same routing/fallback/capacity/audit pipeline as production traffic.

Toggle Web search to ask the gateway to enable provider web search:

  • Chat Completions → web_search_options: {} (OpenAI search-class models, Perplexity sonar models)
  • Responses → tools: [{ type: 'web_search_preview' }]
  • Anthropic Messages → tools: [{ type: 'web_search_20250305' }]

Providers that don't recognise the field ignore it.

Gateway envelope inputs​

A side panel on the page exposes the optional vmx envelope inputs so you can exercise correlation, metadata-tagging, and per-call resource overrides without leaving the playground:

  • Correlation ID — forwarded as vmx.correlationId; lets you group multi-step calls in the audit / usage views.
  • Metadata — key/value pairs forwarded as vmx.metadata. The playground always tags sends with playground=true and the signed-in user_id so audit filters can separate playground traffic from real workload.
  • Resource override JSON — a free-form JSON object merged into the resource via vmx.resourceConfigOverrides for the duration of the call (handy for tweaking model parameters without saving the resource).

Multimodal input​

The chat box accepts:

  • Images — paste from the clipboard (most common: screenshots), drag/drop onto the chat panel, or click the paperclip.
  • Audio — drag/drop or paperclip; encodes inline as base64.
  • Files — generic; same upload flow.

Each attachment shows up as a chip above the input with its type icon, filename, size, and a remove (✕) button. Total attachment payload is capped at 100 MB per request — the BFF and Fastify gateway are both configured for that ceiling.

The gateway forwards the attachment as the right OpenAI content type (image_url data URL, input_audio, file) and lets the provider decide whether it can handle that modality. If the provider rejects it, the error surfaces in the chat panel.

Audit + storage​

Audit rows store metadata only for attachments — the mime type, byte size, and a SHA-256 of the base64 payload — not the bytes themselves. This keeps the request_audit.request_payload JSONB column small and avoids carrying potentially-PII'd bytes around. The live request to the provider always includes the full bytes; only the audit-row copy is sanitised.

See also​