Platform and infrastructure teams
- One endpoint for every pipeline stage
- Residency enforced without forking app code
- Cache behaviour you can inspect across borders
Your documents stay in an in-country store. The model reasoning over them must stay there too. That includes embedding and reranking, the steps teams most often miss.
Embed, rerank, generate
Refuse, never fall back
The problem
A pipeline is audited for where its output runs, then calls a US-hosted embedding endpoint and a global reranker.
Failover is set for availability, so a residency-bound request can land in whatever region answers first.
The same corpus is indexed twice, in two regions, because no one chose the authoritative copy.
The approach
Embeddings, reranking and model output use the same jurisdiction policy as chat. A leaking stage fails closed instead of falling back.
Region is declared per request, key or workspace. A retrieval service pinned to Korea cannot answer from Tokyo.
Each stage records its serving region in an exportable residency record.
Models for this workload
| Model | Provider | Context | Input, per 1M | Regions | Residency |
|---|---|---|---|---|---|
| DeepSeek V4 Pro | DeepSeek | 1M | $0.3 | CN · cn-beijing | In region |
Who this is for
Other solutions
02
Images, audio and video under text's rules
03
Long tool loops that stay inside a jurisdiction
04
Second waveCompletions and refactors without exporting the repository
05
Second waveStreaming assistants with residency that holds under load
06
Second waveRanking and reranking inside the same boundary as model output
Describe a different workload
Start