Chat

The Chat tab is a streaming, multi-turn conversation with a model you pick. It is backed by the chat edge function.

Change the chat system prompt in this project. Edit the chat edge function, and tell me what else reads that prompt.

A chat thread showing a markdown-formatted assistant reply, with the thread's model shown as a badge in the header

Change the prompt or reply length

To change	Edit	Then
The system prompt	`supabase/functions/chat/index.ts`	Redeploy
Reply length	`MAX_OUTPUT_TOKENS` in the same file (currently `2048`)	Redeploy
Which models appear	`config.js`, and the function’s allow-list	See Models
The UI	`src/components/chat/`	Nothing

No system prompt is set by default, so the function forwards the conversation as-is. Add one for your product by passing system to streamText:

supabase/functions/chat/index.ts


const result = streamText({
  model: openrouter.chat(model),
  system: 'You are a helpful assistant for …',
  messages: modelMessages,
  maxOutputTokens: MAX_OUTPUT_TOKENS,
});


supabase functions deploy chat

Raising MAX_OUTPUT_TOKENS costs proportionally more per message.

Gotchas

Keep expo/fetch if you replace the transport, because React Native’s global fetch buffers the whole response body before resolving, so every token arrives at once.

The model is fixed once a thread exists, so a thread’s history stays consistent with the model that produced it. ModelPicker is disabled on existing threads. To switch, start a new chat.

Attachments are resized to a 1024px long edge and re-encoded as JPEG at quality 0.7, because vision models bill by image size. The picker’s own quality option compresses without capping dimensions, so it is not enough on its own.

Error codes

The function answers with a JSON body containing a code:

Code	HTTP	Meaning
`invalid_model`	400	Slug not in the function’s allow-list. See Models.
`model_not_vision`	400	An image was attached to a model that can’t see.
`invalid_file_url`	400	A file part whose URL isn’t a `chat-media` signed URL.
`pro_required`	403	An `imageOutput` model requested by a non-Pro user.
`not_found`	404	The `threadId` doesn’t belong to the caller.
`quota_exceeded`	402	Free allowance used up. See Quota & Paywall.
`provider_not_configured`	500	`OPENROUTER_API_KEY` isn’t set. See Secrets.

Files

Route	File
Thread list	`src/app/(protected)/(tabs)/chats.tsx`
A conversation	`src/app/(protected)/chat/[threadId].tsx`
Transport	`src/lib/ai/chatTransport.ts`
Attachments	`src/lib/storage/chatMedia.ts`
Server	`supabase/functions/chat/index.ts`

How it works

Threads. The client never sends its own thread id. The function creates one on the first message and returns it in an x-thread-id header, titled from the first 40 characters of the message, or New chat if that message was image-only.

Attachments. An upload goes to chat-media/{userId}/<uuid>.jpg as a signed URL, which the server rewrites to the bare storage path before persisting, because signed URLs expire. resolveChatMediaUrl() re-signs the path on demand when history renders. The function only accepts a file part whose URL is a chat-media signed URL, so it cannot be tricked into persisting an arbitrary one.

Server-side inlining. Model providers reject Supabase Storage signed URLs outright (Only HTTPS URLs are supported., Cannot fetch from private/localhost URLs). Before calling the model, the function downloads each image with the service-role client and inlines it as a base64 data: URL, which works with any provider and keeps the bucket private. See inlinePartsForModel() in supabase/functions/_utils/ai.ts.

Image replies. A model flagged imageOutput can answer with images. These models are Pro-only, and their replies upload separately to chat-media/{userId}/gen/ so they survive in history.