Embeddings Image Speech

Operator embeds 5000 chunks by calling `embed({ model, value })` in a Promise.all loop and hits per-request rate limits.

Use `embedMany({ model, values: chunks, maxParallelCalls })` — it batches into provider-allowed group sizes and parallelizes within the per-provider limit. The SDK manages batch boundaries; the integrator does not need a custom chunker. Returns embeddings[] in input order plus aggregate usage.

Operator hand-rolls cosine similarity over embedding[] with a subtle bug (forgot to normalize).

Use the SDK's `cosineSimilarity(a, b)` helper from 'ai'. It handles normalization correctly. Hand-rolled implementations are a frequent source of off-by-norm bugs; prefer the SDK helper for any production retrieval.

Operator stores embeddings from text-embedding-3-small (1536-dim) and later switches to text-embedding-3-large (3072-dim) without re-indexing.

Embedding dimension is model-specific. Store the model id alongside vectors and refuse cross-model similarity. On model swap, re-embed the corpus before switching reads. Use the SDK's per-model dimension info (where available) to validate at write time.