Dataset chats

Chats are project training examples. A chat contains ordered system, user, and assistant messages plus optional name, description, and tags.

CRUD and bulk endpoints

GET    /api/v1/projects/{owner_slug}/{project_slug}/chats
POST   /api/v1/projects/{owner_slug}/{project_slug}/chats
POST   /api/v1/projects/{owner_slug}/{project_slug}/chats/bulk
GET    /api/v1/projects/{owner_slug}/{project_slug}/chats/{chat_id}
PUT    /api/v1/projects/{owner_slug}/{project_slug}/chats/{chat_id}
DELETE /api/v1/projects/{owner_slug}/{project_slug}/chats/{chat_id}
GET    /api/v1/projects/{owner_slug}/{project_slug}/chats/categories
GET    /api/v1/projects/{owner_slug}/{project_slug}/tags

Create one example:

{
  "name": "returns-001",
  "description": "Approved returns answer",
  "tags": ["support"],
  "messages": [
    {"role": "user", "text": "Can I return this?"},
    {"role": "assistant", "text": "Yes, within 30 days."}
  ]
}

List and detail responses include a computed message_count, durable updated_at, and nullable source. Source is csv, api, editor, or playground; old rows can be null. API-key writes are always classified by the server and cannot spoof provenance.

Automatically saved text inferences from the OpenAI-compatible endpoint have source=api and the inference and model:<model reference> tags. They follow the owner's sampling setting and never capture another caller's shared/public traffic. A manual Playground save uses source=playground. Historical missing sources remain null; names and editable tags are not used to guess origin.

Browsing at scale

The list always uses bounded cursor pages, without a mode opt-in. Numbered page and page_size parameters are not part of this contract:

GET .../chats?limit=25
GET .../chats?limit=25&cursor=<next_cursor>
GET .../chats?limit=25&before=<previous_cursor>
GET .../chats?limit=25&boundary=last

cursor, before, and boundary are mutually exclusive and opaque. Cursors belong to this project. Responses include next_cursor, previous_cursor, is_last_page, total_count, total_count_capped, and search_applied. Counts are capped at 10,000 rather than scanning an arbitrarily large dataset.

Filters run before pagination. Search matches chat names and message text; terms shorter than two characters are not applied and terms longer than 128 characters return 422. Use tags, source, and normalized categories filters as published by the endpoint. GET .../chats/categories?q= returns at most 50 visible category facets with counts.

Mutability and training

Editing or deleting a chat changes the current dataset, but never rewrites an already-created training job. Before it starts, every job freezes its selected input and train/evaluation membership. Use the job dataset endpoints when you need the exact historical input rather than the current chat.