Search and relevance teams
- Swap rerankers without a new compliance review
- Compare quality by region
Second wave
Ranking and reranking inside the same boundary as model output
Search quality depends on models that rarely appear in architecture reviews: the embedder, reranker and query rewriter.
- Per-stage
- Policy coverage
- Visible
- Index-boundary drift
- Overview
- Coverage
Written at second wave
01
The problem
Policy does not see the ranking stack
Failure mode 01
Query rewriting and reranking are treated as infrastructure, rather than model calls subject to residency.
Failure mode 02
Global and regional indexes drift, and that drift is rarely intentional.
02
The approach
Ranking-stage requests
Measure 01
Embedding and reranking calls carry the model-output jurisdiction policy.
Measure 02
Region is declared at the index boundary, so drift is visible before discovery.
Measure 03
A refusal is explicit when a ranking stage cannot run in-region.
03
Models for this workload
Starting points and serving regions
| Model | Provider | Context | Input, per 1M | Regions | Residency |
|---|---|---|---|---|---|
| DeepSeek V4 Pro | DeepSeek | 1M | $0.3 | CN · cn-beijing | In region |
04
Who this is for
Audience
Platform teams
- One policy for search and chat
- Refusal telemetry that separates policy from failure
Compliance
- Ranking stages inside the reviewed boundary
- Every regional index documented
05
Other solutions
Other solutions
01
Enterprise RAG
Retrieval for data that must stay in the country
02
Multimodal
Images, audio and video under text's rules
03
Agentic systems
Long tool loops that stay inside a jurisdiction
04
Second waveCode assistance
Completions and refactors without exporting the repository
05
Second waveConversational AI
Streaming assistants with residency that holds under load
Describe a different workload
Start