chore(ms-ai-architect): refresh KB critical-bucket — 15 files [skip-docs]

KB-currency refresh (critical priority, 2026-06-19) via /architect:kb-update.
15 critical cost/arkitektur-filer re-verifisert mot Microsoft Learn (MCP).

Hovedendringer:
- Azure AI Search Serverless (Preview) prismodell (rag-cost, rag-query,
  vector-storage, rag-enterprise-scale)
- Agentic retrieval GA-split (2026-04-01 REST GA, portal preview) + Foundry IQ
- AI gateway multi-provider (Anthropic v2 / Google Vertex), unified model API
  (preview), Foundry-integrasjon (enterprise-arch, ai-services-cost,
  multi-model, cost-tracking)
- FOCUS-skjema + Cost Mgmt -> ADLS Gen2 -> Fabric -> Power BI pipeline
  (cost-management, cost-allocation)
- FinOps hub AI-agent via Azure MCP / Copilot Studio (budget-forecasting)
- OTel trace-based sampling for logs (observability)
- Korreksjoner: semantic-cache score-threshold = AVSTAND (0.15, ikke 0.85
  cosine-likhet); llm-emit-token-metric maks 5 custom dimensions;
  Foundry RBAC-renames (Owner/Contributor = kun mgmt); Researcher/Analyst er
  ikke agenter
- Alle 15 Last updated -> 2026-06-19

High-bucket (49 filer) UTSATT til egen sesjon (operatoer-valg).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01REiKFhP4w6xGXXqWKpPCJJ

skills/ms-ai-security/references/cost-optimization/semantic-caching-patterns.md

@@ -1,6 +1,6 @@
 # Semantic Caching for AI Workloads
 
-**Last updated:** 2026-02
+**Last updated:** 2026-06-19
 **Status:** GA
 **Category:** Cost Optimization & FinOps for AI
 
@@ -65,6 +65,8 @@ Semantic caching består av fire hovedkomponenter:
 - `0.70-0.84`: Liberal matching, høyere cache hit rate men lavere presisjon
 - **Start med 0.85 og juster basert på cache hit rate og user feedback** (MEDIUM confidence)
 
+**Viktig — `score-threshold` i APIM-policyen er en AVSTAND, ikke cosine-likhet (MCP 2026-06):** Tabellen over gjelder generell cosine-**likhet** (høyere = strengere), som brukes av application-level vector-stores. Men APIM-policyene `llm-semantic-cache-lookup` / `azure-openai-semantic-cache-lookup` bruker et `score-threshold` som er en semantisk **avstand**: prompts med score *over* terskelen bruker IKKE cachen, så **lavere terskel = strengere matching**. Microsofts eget eksempel bruker `score-threshold="0.15"`. Sett derfor en LAV verdi (~0.1–0.2) i APIM-policyene under — IKKE 0.85.
+
 ---
 
 ## Arkitekturmønstre
@@ -88,7 +90,7 @@ Client → APIM (semantic cache policies) → Azure Managed Redis (RediSearch)
 ```xml
 <!-- Inbound: Cache Lookup -->
 <azure-openai-semantic-cache-lookup
-    score-threshold="0.85"
+    score-threshold="0.15"
     embeddings-backend-id="embeddings-backend"
     embeddings-backend-auth="system-assigned"
     ignore-system-messages="true"
@@ -313,7 +315,7 @@ L3: Azure OpenAI → 500-2000ms latency
 
 | Parameter | Verdi | Forklaring |
 |-----------|-------|------------|
-| `score-threshold` | 0.85 (anbefalt) | Minimum similarity for cache hit (0-1) |
+| `score-threshold` | 0.15 (lav!) | Semantisk **avstand**-terskel: prompts over terskelen bruker IKKE cachen → **lavere = strengere**. IKKE cosine-likhet. MS-eksempel: 0.15 |
 | `embeddings-backend-id` | "embeddings-backend" | Backend ID for embeddings deployment |
 | `embeddings-backend-auth` | "system-assigned" | Bruker APIM managed identity |
 | `ignore-system-messages` | true | Ignorer system messages i similarity-beregning |
@@ -398,7 +400,7 @@ Azure AI Foundry models (via Model Inference API) støttes med generic LLM polic
 ```xml
 <!-- Use llm-semantic-cache-lookup instead of azure-openai-semantic-cache-lookup -->
 <llm-semantic-cache-lookup
-    score-threshold="0.85"
+    score-threshold="0.15"
     embeddings-backend-id="embeddings-backend"
     embeddings-backend-auth="system-assigned">
     <vary-by>@(context.Subscription.Id)</vary-by>