Skip to content

What's new in Kumiko AI

Enterprise

This page tracks the user-visible Enterprise AI surface. It is intentionally organized around workflows rather than model release names, so the architecture remains useful as providers change.

  • In-app assistant: search and read live records, ask clarifying questions, propose writes, review them in the normal form, and enforce tenant caps and approval rules.
  • Feature builder: generate a feature, suggest focused patches, review the source, validate it, and publish a versioned pattern.
  • Document AI: extract schema-shaped data from text or PDFs, classify messages, and tune each step through a tenant-editable pipeline policy.
  • Assistant attachments: attach PDF, image, CSV, TXT, and Word files to a conversation. Word files are read as extracted text by a document-ingest parser provider. Users can also paste or drop files and take a photo with the camera.
  • Assistant window and conversations: dock the assistant, float it as a movable panel, minimize it to a bubble, or pop it out into its own window. Earlier conversations are listed, reopened, and renamed.
  • Voice: push-to-talk, live transcription while speaking, read-aloud for answers, and a hands-free conversation loop, metered in audio seconds and characters.
  • Template editing: ask the assistant to change a mail or PDF template in plain text, for example “put the logo in the header” or “change the signup mail”. It shows the planned changes as a preview and saves them as a draft only after you confirm. Publishing is a separate action that needs its own confirmation. A template can also be previewed with a real record instead of sample values: search for a record, pick it, and the placeholders are filled from it. See Design mail and PDF templates.
  • Prompt and policy operations: edit prompts, draft changes, activate them, run golden fixtures, and roll back.
  • Provider choice: tenants add Anthropic, OpenAI and OpenAI-compatible connections (OpenRouter and local servers included) and pick standard models per job. OpenAI and OpenAI-compatible connections also handle transcription and read-aloud. Deterministic mocks cover tests.
  • EU only and zero data retention: a connection can be bound to an EU OpenRouter endpoint with zdr and data_collection: "deny" forced, and the installation can enforce that for all connections.
  • Enterprise 0.65.0 removes the legacy provider ids. The old ids anthropic, openai-compat and openrouter are no longer migrated on boot, and a prompt that still stores one of them counts as an unknown connection. Apps with ai_connections run kumiko migrate generate during the bump and uncomment the two generated lines that drop the index ai_connections_tenant_legacy_provider_uniq and the column legacy_provider_id. Enterprise ships no SQL of its own, and rebuilds need no upcaster.

When a new AI workflow becomes available, update this page with the user problem it solves, its runnable example, and the relevant governance controls.