Migrating Messages to AI SDK v5 UIMessage Format

[khanhduyvt0101]

Problem

After upgrading to AI SDK v5, I need to migrate existing chat messages in PostgreSQL. My messages currently have both old format fields (content, annotations) and new parts array, but tool invocations still use the old state: "result" instead of v5's 4-state system. I also store messages in jsonb type.

Current Message Format (Mixed)

{
  "id": "msg-S8wpAp3QWIHXEuufOxeA2vEt",
  "role": "assistant",
  "content": "can you find papers...",  // Old field - should be removed
  "annotations": [{"type": "file", "hasFile": false}],  // Old field
  "parts": [
    {
      "type": "tool-invocation",
      "toolInvocation": {
        "state": "result",  // Old state - should be "output-available"
        "result": { "results": [...] },
        "toolName": "semantic_scholar_search"
      }
    },
    {
      "type": "text", 
      "text": "I found several papers..."
    }
  ]
}

Target AI SDK v5 Format

interface UIMessage {
  id: string;
  role: "user" | "assistant" | "system";
  parts: MessagePart[];
  metadata?: Record<string, any>;
}

// Tool states: "input-streaming" | "input-available" | "output-available" | "output-error"

Questions

1. Best migration approach for ~500k messages?

   • Batch migration script vs on-the-fly conversion?
   • How to handle the state mapping (result → output-available)?

2. Should I remove legacy fields (content, annotations) or keep for backward compatibility?

3. How to handle tool invocations that might fail during migration?

Looking for advice on:

• PostgreSQL JSONB migration best practices
• Handling large-scale migrations without downtime
• Rollback strategies

Thanks! 🙏

[Theonlyhamstertoh]

Same issue here, any update on this?