> embedding · dim=1536 · cosine=0.94
> inference stream · token 128/512 · latency 42ms
> neural.attention · layer 12/24 · heads=8
> RAG.retrieve({ locale: 'zh-HK', top_k: 3 })
> agent.tool_call('members.list') → ok
> model.route({ tenant, context: 'Victoria Harbour' })
> vector.index.upsert · batch 64 · HK region
> stream.complete · finish_reason=stop