Decision D21 · 21 of 21
LLM provider and data residency (new)
One gateway, low-cost default model, no plant data to a provider the customer has not approved
RECOMMENDED
- Status
- RECOMMENDED — awaiting Vinayak
- Recommendation
- All LLM calls go through one gateway in KR with a per-tenant monthly token budget, prompt caching, and a provider allow-list per customer. Default model: a low-cost model from a provider with a no-training, zero-retention API contract. Plant data goes only to providers the customer has approved in writing; a customer that approves none gets Ask with deterministic tools only.
- Why
- LLM cost per plant is small next to infrastructure, but data residency is a procurement blocker for Indian manufacturers. A gateway makes the provider a setting, not a rewrite.
- Evidence
- Provider pricing pages (DeepSeek, OpenAI, Anthropic, checked 7 Oct 2026); section 8 agent rules (read-only tools, evidence IDs).
- Reversible
- Yes. Provider is configuration.
- Blocks
- Ask in production, running-cost LLM line.
Page history: last 2 changes
- docs(research): retire stale research to archive/research-2026-10 with a register
ab84821 - docs(technical): split decision board into DECISIONS.md
b4db9d4