Chat and Summaries

Model

Chat, summaries and prompt enhancement use TEXT_MODEL in packages/ai/src/text.ts, currently @cf/google/gemma-4-26b-a4b-it on Workers AI through the AI binding. No model key is needed. In mock mode, mockText returns a deterministic answer so the chat works offline.

Changing the model means checking its input format, streaming output, context length, cost and the tests together; the template is not a gateway that configures every provider automatically.

Chat

POST /api/chat:

{
  "prompt": "Write three short taglines for an AI image app",
  "requestId": "e60b2e4b-32b1-42a5-bfd0-dbb20ecb37e9",
  "history": [{ "prompt": "Earlier question", "answer": "Earlier answer" }]
}

• prompt is 1 to 16,000 characters. requestId is optional; when present it must be a UUID, created per message.
• history holds up to 8 recent turns (4,000 characters per prompt, 8,000 per answer) so the model keeps context. The studio at /app/ai sends it automatically.
• The response is text/event-stream. Read it chunk by chunk and handle a broken stream; do not call response.json(). The studio renders Markdown as it arrives.

Each message costs quotas.chatCredits (1). The server checks the balance and writes a generation row that reserves the credit before calling the model. A finished answer completes the generation; a failed or broken stream releases the credit once.

Summaries

1. Upload your own text/plain or text/markdown file at /app/files (up to 1 MiB).
2. Call POST /api/summary with { "fileId": "your-owned-file-uuid" }.
3. The response is the scheduled task, not the summary. The job runs on the JOBS queue; follow its state in the workbench.

A summary costs quotas.summaryCredits (5) and reads the first 16,000 characters. It does not split long documents, OCR images or parse arbitrary PDFs.

Verify

• Empty or overlong prompts, too few credits and files owned by someone else all fail without a charge.
• A normal chat streams and leaves one ledger line; an interrupted stream leaves the balance unchanged.
• A summary succeeds once even when the queue redelivers the message.

Continue with Credits and cost control and Testing.