feat(ai-sdk): upgrade to AI SDK v6 with embedding model support

[ItsWendell]

Summary

This PR upgrades the @temporalio/ai-sdk package to support AI SDK v6 and adds embedding model support for building RAG pipelines in Temporal workflows.

Key Changes

• AI SDK v6 Upgrade: Migrate from AI SDK v5 (@ai-sdk/provider v2 types) to AI SDK v6 (@ai-sdk/provider v3 types)
• Embedding Model Support: Add TemporalEmbeddingModel class and invokeEmbeddingModel activity for durable embedding generation in workflows
• Enhanced Provider Options: New TemporalProviderOptions interface allows configuring separate ActivityOptions for language vs embedding models
• Improved MCP Types: Use jsonSchema() helper instead of manual symbol hacks, proper FlexibleSchema typing
• Updated Tests: All AI SDK tests updated for v6 compatibility with new embedding model tests

Motivation

The current ai-sdk package uses AI SDK v5 and lacks embedding model support. This limits its use in modern AI applications that require:

1. Embedding models - Essential for RAG pipelines, semantic search, and vector-based retrieval
2. AI SDK v6 - Latest features and improvements from the Vercel AI SDK
3. Per-model configuration - Different timeout requirements for fast language model calls vs potentially slower embedding batches

Breaking Changes

This is a breaking change that requires updating peer dependencies:

Dependency	Before	After
@ai-sdk/provider	^2.0.0	^3.0.0
ai	^5.0.91	^6.0.0
@ai-sdk/mcp	^0.0.8	^1.0.0

Usage Examples

Embedding Models in Workflows

import { embedMany } from 'ai';
import { temporalProvider } from '@temporalio/ai-sdk';

export async function ragWorkflow(documents: string[]) {
  // Generate embeddings through Temporal activities
  const { embeddings } = await embedMany({
    model: temporalProvider.embeddingModel('text-embedding-3-small'),
    values: documents,
  });
  
  // Use embeddings for semantic search...
  return embeddings;
}

Per-Model Activity Options

import { TemporalProvider } from '@temporalio/ai-sdk';

const provider = new TemporalProvider({
  default: {
    startToCloseTimeout: '5 minutes',
    retry: { maximumAttempts: 3 },
  },
  languageModel: {
    startToCloseTimeout: '10 minutes', // LLM calls can take longer
  },
  embeddingModel: {
    startToCloseTimeout: '2 minutes', // Embeddings are faster
  },
});

Test Results ✅

All AI SDK tests pass:

✔ Hello world agent responds in haikus (704ms)
✔ Middleware (530ms)
✔ Generate object (551ms)
✔ Embedding model generates embeddings (511ms)
✔ Tools workflow can use AI tools (591ms)
✔ Telemetry (1.3s)
✔ MCP Use (8.5s) - passes locally, skipped in CI

6 tests passed
1 test skipped (MCP - skipped in CI due to server setup)

Test Plan

• Update TestModel and TestProvider to v3 interfaces
• Add TestEmbeddingModel for mock embedding responses
• Add embedding workflow test (embeddingWorkflow)
• Verify invokeEmbeddingModel activity scheduling
• Update tool definition syntax (inputSchema instead of parameters)
• Update middleware type to LanguageModelV3Middleware
• Verify MCP test passes locally with jsonSchema() helper
• All tests pass locally

[CLAassistant]

All committers have signed the CLA.