diff --git a/scripts/kb-eval/data/spor0-fix-manifest.json b/scripts/kb-eval/data/spor0-fix-manifest.json index 52f3eaf..4b7febd 100644 --- a/scripts/kb-eval/data/spor0-fix-manifest.json +++ b/scripts/kb-eval/data/spor0-fix-manifest.json @@ -4,7 +4,10 @@ "derived_from": "gold-correctness-set.json (frozen) + judge-bakeoff-results-v2.json", "count": 38, "files_touched": 25, - "note": "This is a SAMPLE of corpus errors (255 volatile claims). Spor 1 surfaces the full set corpus-wide." + "note": "This is a SAMPLE of corpus errors (255 volatile claims). Spor 1 surfaces the full set corpus-wide. [2026-06-29] Spor 0 executed: 37 fixes applied (each value live-verified per source before edit; recurrences of each wrong fact fixed file-wide), 1 rejected (#8 model-router-GA: file already correct). Adjacent (non-38) findings captured for Spor 1.", + "applied": 37, + "rejected": 1, + "applied_date": "2026-06-29" }, "fixes": [ { @@ -12,494 +15,609 @@ "file": "skills/ms-ai-advisor/references/platforms/azure-ai-foundry.md", "skill": "ms-ai-advisor", "claim_type": "version", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "DeepSeek-R1-0528 Reasoning 131072 tokens", "correction": "Fil 131072 vs kilde 163840. Trolig forvekslet med V3-familien.", "source": "https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-advisor/platforms/azure-ai-foundry.md#7", "file": "skills/ms-ai-advisor/references/platforms/azure-ai-foundry.md", "skill": "ms-ai-advisor", "claim_type": "version", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "computer-use-preview (2025-03-11) via Responses API; 8192 kontekst, 1024 output; aka.ms/oai/cuaaccess", "correction": "Computer Use bruker nå gpt-5.4-modell med computer-tool; URL og 8192/1024 avløst.", "source": "https://learn.microsoft.com/azure/foundry-classic/openai/how-to/computer-use", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-advisor/platforms/model-catalog-2026.md#4", "file": "skills/ms-ai-advisor/references/platforms/model-catalog-2026.md", "skill": "ms-ai-advisor", "claim_type": "tpm", - "verdict": "wrong", + "verdict": "confirm-fix", "wrong_assertion": "DeepSeek-R1/V3 input/output-ratio 1 (standard, som input)", "correction": "Fil: DeepSeek output-to-input-ratio = 1. Kilde: 4. Motsagt av kilden.", "source": "https://learn.microsoft.com/azure/foundry/openai/how-to/provisioned-throughput-sizing", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-advisor/platforms/model-catalog-2026.md#6", "file": "skills/ms-ai-advisor/references/platforms/model-catalog-2026.md", "skill": "ms-ai-advisor", "claim_type": "region", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "gpt-5.5 er eneste modell med Data Zone Standard (EU-residens) i Norway East", "correction": "gpt-5.4 (2026-03-05) har OGSÅ Data Zone Standard i norwayeast. Tidsdrift.", "source": "https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure-region-availability", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-advisor/platforms/model-catalog-2026.md#7", "file": "skills/ms-ai-advisor/references/platforms/model-catalog-2026.md", "skill": "ms-ai-advisor", "claim_type": "status", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "gpt-5/gpt-5-codex/gpt-5-pro krever registrering aka.ms/oai/gpt5access", "correction": "Kilde: tilgang ikke lenger begrenset. Tidsdrift.", "source": "https://learn.microsoft.com/azure/foundry/openai/how-to/reasoning", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-advisor/prompt-engineering/chain-of-thought-prompting.md#3", "file": "skills/ms-ai-advisor/references/prompt-engineering/chain-of-thought-prompting.md", "skill": "ms-ai-advisor", "claim_type": "taxonomy", - "verdict": "wrong", + "verdict": "confirm-fix", "wrong_assertion": "reasoning_summary for GPT-5-serien støtter auto/concise/detailed", "correction": "Kilden: GPT-5-serien støtter IKKE concise. Påstanden feil.", "source": "https://learn.microsoft.com/azure/foundry/openai/how-to/reasoning", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-advisor/prompt-engineering/real-time-reasoning-performance.md#5", "file": "skills/ms-ai-advisor/references/prompt-engineering/real-time-reasoning-performance.md", "skill": "ms-ai-advisor", "claim_type": "region", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Realtime API regions: East US 2, Sweden Central (global)", "correction": "Kilde: tilgjengelig for global deployments, ikke begrenset til de to. To-region-begrensning avløst.", "source": "https://learn.microsoft.com/azure/foundry/openai/how-to/realtime-audio", "v2_judge_caught": false, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-advisor/prompt-engineering/real-time-reasoning-performance.md#6", "file": "skills/ms-ai-advisor/references/prompt-engineering/real-time-reasoning-performance.md", "skill": "ms-ai-advisor", "claim_type": "status", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Realtime API fortsatt public preview (jan 2026); ikke SLA", "correction": "Kilde: GA-endepunkt + GA-modeller. Kroppens preview-status foreldet (selvmotsigende med filens header).", "source": "https://learn.microsoft.com/azure/foundry/openai/how-to/realtime-audio", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-advisor/prompt-engineering/reasoning-models-o1-o3-optimization.md#7", "file": "skills/ms-ai-advisor/references/prompt-engineering/reasoning-models-o1-o3-optimization.md", "skill": "ms-ai-advisor", "claim_type": "status", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Limited access for o3-pro/gpt-5-pro/gpt-5-codex via aka.ms-lenker", "correction": "Kilde: tilgang ikke lenger begrenset. Tidsdrift.", "source": "https://learn.microsoft.com/azure/foundry/openai/how-to/reasoning", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-engineering/azure-ai-services/ai-services-cost-optimization.md#5", "file": "skills/ms-ai-engineering/references/azure-ai-services/ai-services-cost-optimization.md", "skill": "ms-ai-engineering", "claim_type": "taxonomy", - "verdict": "wrong", + "verdict": "confirm-fix", "wrong_assertion": "Reservasjonsrabatt 1-års eller 3-års Azure Reservations", "correction": "Kilde: 1-måned eller 1-år. Ingen 3-års term.", "source": "https://learn.microsoft.com/azure/foundry/openai/concepts/provisioned-throughput-billing", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#3", "file": "skills/ms-ai-engineering/references/azure-ai-services/ai-services-vs-foundry-tools-selection.md", "skill": "ms-ai-engineering", "claim_type": "version", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Modellserie (2026-02): o4-mini topp-reasoning, o3, GPT-4o, GPT-3.5-Turbo, DALL-E 3, Whisper", "correction": "Lineup headet av GPT-5.5/5.4/...+GPT-4.1; GPT-5.x/GPT-4.1 mangler, GPT-3.5-Turbo ikke lenger i highlights.", "source": "https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-engineering/azure-ai-services/ai-services-vs-foundry-tools-selection.md#5", "file": "skills/ms-ai-engineering/references/azure-ai-services/ai-services-vs-foundry-tools-selection.md", "skill": "ms-ai-engineering", "claim_type": "sku", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Model Catalog 100+ modeller", "correction": "Kilde over 1900. '100+' grovt understated.", "source": "https://learn.microsoft.com/azure/foundry/what-is-foundry", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-engineering/mlops-genaiops/genaiops-llm-specific-practices.md#2", "file": "skills/ms-ai-engineering/references/mlops-genaiops/genaiops-llm-specific-practices.md", "skill": "ms-ai-engineering", "claim_type": "sku", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Model Catalog 1600+ foundation models", "correction": "Kilde over 1900. 1600+ utdatert lavere tall.", "source": "https://learn.microsoft.com/azure/foundry/concepts/foundry-models-overview", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#3", "file": "skills/ms-ai-engineering/references/mlops-genaiops/llm-evaluation-production.md", "skill": "ms-ai-engineering", "claim_type": "sku", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "MLflow built-in judges inkluderer også Completeness, Fluency, Equivalence", "correction": "Gjeldende Databricks-liste inneholder IKKE disse tre.", "source": "https://learn.microsoft.com/azure/databricks/mlflow3/genai/eval-monitor/concepts/judges/", "v2_judge_caught": false, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-engineering/mlops-genaiops/llm-evaluation-production.md#7", "file": "skills/ms-ai-engineering/references/mlops-genaiops/llm-evaluation-production.md", "skill": "ms-ai-engineering", "claim_type": "version", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Azure AI Evaluation SDK v1.14.0", "correction": "Gjeldende v1.17.0.", "source": "https://learn.microsoft.com/python/api/overview/azure/ai-evaluation-readme", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-engineering/rag-architecture/embedding-models-selection.md#6", "file": "skills/ms-ai-engineering/references/rag-architecture/embedding-models-selection.md", "skill": "ms-ai-engineering", "claim_type": "sku", - "verdict": "wrong", + "verdict": "confirm-fix", "wrong_assertion": "Azure AI Search Basic tier 1 GB storage", "correction": "Kilde: Basic 15 GB (post-april) eller 2 GB (eldre). 1 GB var aldri Basic. Inkonsistent med rag-cost-optimization.md.", "source": "https://learn.microsoft.com/azure/search/search-limits-quotas-capacity", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-engineering/rag-architecture/embedding-models-selection.md#7", "file": "skills/ms-ai-engineering/references/rag-architecture/embedding-models-selection.md", "skill": "ms-ai-engineering", "claim_type": "sku", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Standard S1 25 GB storage", "correction": "S1 nå 160 GB per partisjon (nye services).", "source": "https://learn.microsoft.com/azure/search/search-limits-quotas-capacity", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-engineering/rag-architecture/rag-context-windows.md#2", "file": "skills/ms-ai-engineering/references/rag-architecture/rag-context-windows.md", "skill": "ms-ai-engineering", "claim_type": "version", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "GPT-4.1 series opp til 200k+ context window", "correction": "Kilde ~1M. '200k+' understater grovt.", "source": "https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "v2_judge_caught": false, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-engineering/rag-architecture/rag-context-windows.md#3", "file": "skills/ms-ai-engineering/references/rag-architecture/rag-context-windows.md", "skill": "ms-ai-engineering", "claim_type": "version", - "verdict": "wrong", + "verdict": "confirm-fix", "wrong_assertion": "o3-mini 128k context window", "correction": "Kilde: 200k input. o3-mini har 200k.", "source": "https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-engineering/rag-architecture/rag-context-windows.md#5", "file": "skills/ms-ai-engineering/references/rag-architecture/rag-context-windows.md", "skill": "ms-ai-engineering", "claim_type": "taxonomy", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "TPM-tabell med Default tier / Enterprise tier-kolonner", "correction": "Default/Enterprise erstattet av Quota Tiers (Tier 0-6).", "source": "https://learn.microsoft.com/azure/foundry/openai/quotas-limits", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-engineering/rag-architecture/rag-cost-optimization.md#5", "file": "skills/ms-ai-engineering/references/rag-architecture/rag-cost-optimization.md", "skill": "ms-ai-engineering", "claim_type": "sku", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Per-partisjon storage S1 25/S2 100/S3 200/L1 1TB/L2 2TB", "correction": "Fil-verdier pre-april-2024. Nye: S1 160/S2 512/S3 1024/L1 2048/L2 4096. Internt inkonsistent (Basic oppdatert, øvrige ikke).", "source": "https://learn.microsoft.com/azure/search/search-limits-quotas-capacity", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-engineering/rag-architecture/rag-cost-optimization.md#6", "file": "skills/ms-ai-engineering/references/rag-architecture/rag-cost-optimization.md", "skill": "ms-ai-engineering", "claim_type": "sku", - "verdict": "wrong", + "verdict": "confirm-fix", "wrong_assertion": "Search Units L1 1-12, L2 1-12", "correction": "Kilde: maks 36 SU for L1/L2. Forveksler partisjonstall med SU-maks.", "source": "https://learn.microsoft.com/azure/search/search-limits-quotas-capacity", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-governance/monitoring-observability/endpoint-health-and-capacity-planning.md#2", "file": "skills/ms-ai-governance/references/monitoring-observability/endpoint-health-and-capacity-planning.md", "skill": "ms-ai-governance", "claim_type": "taxonomy", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Metrikker PTUUtilization, ProcessingTime, TokensGenerated", "correction": "Finnes ikke som REST-navn. Faktisk: AzureOpenAIProvisionedManagedUtilizationV2, AzureOpenAITimeToResponse, GeneratedTokens. Metrikknavn drift.", "source": "https://learn.microsoft.com/azure/azure-monitor/reference/supported-metrics/microsoft-cognitiveservices-accounts-metrics", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-governance/monitoring-observability/endpoint-health-and-capacity-planning.md#3", "file": "skills/ms-ai-governance/references/monitoring-observability/endpoint-health-and-capacity-planning.md", "skill": "ms-ai-governance", "claim_type": "tpm", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "RPM/TPM per 1 Unit Capacity: eldre chat 6/1000; o1 1/6000; o3 1/1000; o3-mini 1/10000", "correction": "1 Unit Capacity-modellen erstattet av Quota Tiers (Tier 0-6). Ratioene holder, rammeverket avløst.", "source": "https://learn.microsoft.com/azure/foundry/openai/quotas-limits", "v2_judge_caught": false, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-governance/monitoring-observability/token-usage-tracking-attribution.md#3", "file": "skills/ms-ai-governance/references/monitoring-observability/token-usage-tracking-attribution.md", "skill": "ms-ai-governance", "claim_type": "taxonomy", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Metrikker PromptTokens (input) og CompletionTokens (output)", "correction": "Finnes ikke som REST-navn. Faktisk: ProcessedPromptTokens/InputTokens og GeneratedTokens/OutputTokens. Drift.", "source": "https://learn.microsoft.com/azure/azure-monitor/reference/supported-metrics/microsoft-cognitiveservices-accounts-metrics", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-governance/responsible-ai/responsible-ai-training-awareness.md#2", "file": "skills/ms-ai-governance/references/responsible-ai/responsible-ai-training-awareness.md", "skill": "ms-ai-governance", "claim_type": "status", - "verdict": "wrong", + "verdict": "confirm-fix", "wrong_assertion": "AI-cert (AI-900, AI-102) har ikke formell utløpsdato", "correction": "AI-102 har Renewal Frequency 12 mnd OG retires 2026-06-30. Kun AI-900 utløper ikke; å lumpe AI-102 inn er feil.", "source": "https://learn.microsoft.com/credentials/certifications/azure-ai-engineer", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-governance/responsible-ai/stakeholder-communication-ai-decisions.md#3", "file": "skills/ms-ai-governance/references/responsible-ai/stakeholder-communication-ai-decisions.md", "skill": "ms-ai-governance", "claim_type": "taxonomy", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Copilot Studio inkluderer 25000 messages", "correction": "Kvantum 25000 korrekt, men enheten messages avløst av Copilot Credits (2025-09-01).", "source": "https://learn.microsoft.com/microsoft-365/copilot/pay-as-you-go/copilot-capacity-packs", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-infrastructure/bcdr/ai-foundry-disaster-recovery-planning.md#8", "file": "skills/ms-ai-infrastructure/references/bcdr/ai-foundry-disaster-recovery-planning.md", "skill": "ms-ai-infrastructure", "claim_type": "taxonomy", - "verdict": "wrong", + "verdict": "confirm-fix", "wrong_assertion": "Microsoft Foundry tidligere Azure AI Studio / Azure Machine Learning", "correction": "Brand-evolusjon: Azure AI Studio -> Azure AI Foundry -> Microsoft Foundry. Azure ML er IKKE tidligere navn (egen tjeneste). Konflatering feil.", "source": "https://learn.microsoft.com/azure/foundry/what-is-foundry", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-infrastructure/bcdr/capacity-planning-dr-configurations.md#3", "file": "skills/ms-ai-infrastructure/references/bcdr/capacity-planning-dr-configurations.md", "skill": "ms-ai-infrastructure", "claim_type": "status", - "verdict": "wrong", + "verdict": "confirm-fix", "wrong_assertion": "AI Search 2 vs 3 replikaer: 99.9% vs 99.99% SLA", "correction": "AI Search SLA er 99.9%, ingen 99.99%-nivå. 2 vs 3 = lese vs lese-skrive, ikke 99.9 vs 99.99. Aldri korrekt.", "source": "https://learn.microsoft.com/azure/reliability/reliability-ai-search", "v2_judge_caught": false, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-infrastructure/bcdr/network-resilience-patterns-ai.md#1", "file": "skills/ms-ai-infrastructure/references/bcdr/network-resilience-patterns-ai.md", "skill": "ms-ai-infrastructure", "claim_type": "taxonomy", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Azure Front Door DDoS Protection Standard", "correction": "Gjeldende navn: DDoS Network Protection / DDoS IP Protection. 'Standard'-navnet avløst.", "source": "https://learn.microsoft.com/azure/ddos-protection/ddos-protection-overview", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-security/cost-optimization/azure-ai-foundry-cost-governance.md#6", "file": "skills/ms-ai-security/references/cost-optimization/azure-ai-foundry-cost-governance.md", "skill": "ms-ai-security", "claim_type": "tpm", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Model Quota (TPM) varierer per tier 150K-30M", "correction": "Nå Quota Tiers (Free + Tier 1-6) opp til 225M. 150K-30M foreldet.", "source": "https://learn.microsoft.com/azure/foundry/openai/quotas-limits", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-security/cost-optimization/model-selection-price-performance.md#8", "file": "skills/ms-ai-security/references/cost-optimization/model-selection-price-performance.md", "skill": "ms-ai-security", "claim_type": "status", - "verdict": "outdated", + "verdict": "rejected-file-correct", "wrong_assertion": "Model Router er GA-funksjonalitet i Microsoft Foundry", "correction": "Kilde lenker 'model router for Foundry (preview)'. Fortsatt preview, ikke GA.", "source": "https://learn.microsoft.com/azure/foundry/foundry-models/how-to/model-choice-guide", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": false, + "applied": false, + "verified": "2026-06-29", + "verified_by": "verify-singletons-c (live)", + "note": "Manifest correction WRONG. Model Router reached GA (version 2025-11-18, per whats-new-model-router). Cited model-choice-guide page carries a stale \"(preview)\" label that misled the original subagent. File already correct — NOT changed." }, { "id": "ms-ai-security/cost-optimization/model-selection-price-performance.md#9", "file": "skills/ms-ai-security/references/cost-optimization/model-selection-price-performance.md", "skill": "ms-ai-security", "claim_type": "version", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "AI Builder default GPT-4o-mini for generative (per des 2024)", "correction": "Default nå GPT-4.1 mini; GPT-4o/4o-mini kun US government. Utdatert.", "source": "https://learn.microsoft.com/microsoft-copilot-studio/prompt-model-settings", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-security/cost-optimization/multi-model-strategy-costs.md#2", "file": "skills/ms-ai-security/references/cost-optimization/multi-model-strategy-costs.md", "skill": "ms-ai-security", "claim_type": "taxonomy", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "opptil 18 underliggende modeller", "correction": "2025-11-18 lister nå ~28 modeller. 18 foreldet (in-place oppdatering).", "source": "https://learn.microsoft.com/azure/foundry/openai/concepts/model-router", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-security/cost-optimization/multi-model-strategy-costs.md#3", "file": "skills/ms-ai-security/references/cost-optimization/multi-model-strategy-costs.md", "skill": "ms-ai-security", "claim_type": "tpm", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Rate limits Default/Enterprise GlobalStandard 250/250000, Enterprise 400/400000; DataZone 150/150000", "correction": "Nå tier-basert (Tier 1-6). Default/Enterprise-struktur erstattet av Quota Tiers.", "source": "https://learn.microsoft.com/azure/foundry/openai/concepts/model-router", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-security/cost-optimization/rag-query-cost-reduction.md#2", "file": "skills/ms-ai-security/references/cost-optimization/rag-query-cost-reduction.md", "skill": "ms-ai-security", "claim_type": "sku", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "AI Search tier storage Basic 2/S1 25/S2 100/S3 200 GB per partisjon", "correction": "Pre-april-2024-tall. Nye: Basic 15/S1 160/S2 512/S3 1024.", "source": "https://learn.microsoft.com/azure/search/search-limits-quotas-capacity", "v2_judge_caught": false, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-security/cost-optimization/reserved-capacity-planning.md#6", "file": "skills/ms-ai-security/references/cost-optimization/reserved-capacity-planning.md", "skill": "ms-ai-security", "claim_type": "status", - "verdict": "wrong", + "verdict": "confirm-fix", "wrong_assertion": "Autorenew ON som standard for nye reservasjoner (etter 2025-Q4)", "correction": "Kilde: auto-renewal opt-in; for replacement auto-renew av som standard. Motsier 'ON som standard'.", "source": "https://learn.microsoft.com/azure/cost-management-billing/reservations/azure-openai", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" }, { "id": "ms-ai-security/cost-optimization/vector-storage-cost-optimization.md#5", "file": "skills/ms-ai-security/references/cost-optimization/vector-storage-cost-optimization.md", "skill": "ms-ai-security", "claim_type": "tpm", - "verdict": "outdated", + "verdict": "confirm-fix", "wrong_assertion": "Vector quota (post-April 2024) Basic 1/S1 12/S2 36/S3 72 GB", "correction": "Post-April-2024 vektorkvote er 5/35/150/300 GB. Fil-tall tilsvarer eldre juli-2023-periode.", "source": "https://learn.microsoft.com/azure/search/search-limits-quotas-capacity", "v2_judge_caught": true, "reverify_required": true, - "fixed": false + "fixed": true, + "applied": true, + "verified": "2026-06-29", + "verified_by": "per-source verify-agent (Opus xhigh, live microsoft_docs_fetch) + main-context apply" } ] } diff --git a/skills/ms-ai-advisor/references/platforms/azure-ai-foundry.md b/skills/ms-ai-advisor/references/platforms/azure-ai-foundry.md index bd57a86..306908d 100644 --- a/skills/ms-ai-advisor/references/platforms/azure-ai-foundry.md +++ b/skills/ms-ai-advisor/references/platforms/azure-ai-foundry.md @@ -63,7 +63,7 @@ Katalog med 1,900+ «Foundry Models sold by Azure» (Azure Direct Models), pluss | `gpt-5.2` (2025-12-11) | 400K (input 272K) | 128K | + `gpt-5.2-codex`, `gpt-5.2-chat` (preview) | | `gpt-5.3-codex` (2026-02-24) | 400K (input 272K) | 128K | Kun `-codex`/`-chat`-varianter (ingen bar `gpt-5.3`) | | `gpt-5.4` (2026-03-05) | ~1,05M | 128K | Kontekstsprang til ~1M; + `-mini`/`-nano`/`-pro` | -| `gpt-5.5` (2026-04-24) | ~1,05M (input 922K) | 128K | Eneste GPT-5 med Data Zone (EU) i Norway East | +| `gpt-5.5` (2026-04-24) | ~1,05M (input 922K) | 128K | Data Zone (EU) i Norway East (sammen med gpt-5.4) | GPT-5 støtter reasoning, Chat Completions API, Responses API, structured outputs, text/image input, parallel tool calling. @@ -86,7 +86,7 @@ Alle GPT-4.1-modeller tilgjengelige i Norway East **via Global** (Standard/Provi | DeepSeek-R1 | Reasoning | 163,840 tokens | | DeepSeek-V3 (Legacy) | MoE | 131,072 tokens | | DeepSeek-V3-0324 | MoE | 131,072 tokens | -| DeepSeek-R1-0528 | Reasoning | 131,072 tokens | +| DeepSeek-R1-0528 | Reasoning | 163,840 tokens | | DeepSeek-V3.1 | MoE | 131,072 tokens | DeepSeek-modeller er tilgjengelige i Norway East. @@ -141,16 +141,16 @@ Fullverdig on-device AI inference: - Begrensning: Ikke for distribuert/produksjons-/multi-machine-deployment ### 5. Computer-Using Agents (CUA) -**Status:** Preview (registrering påkrevd) +**Status:** `gpt-5.4` med innebygd computer-tool (limited access — registrering påkrevd) -`computer-use-preview`-modellen (2025-03-11) via Responses API: +`gpt-5.4`-modellen med innebygd `computer`-tool via Responses API (`tools=[{"type": "computer"}]`): - Autonom navigasjon: klikker knapper, fyller skjemaer, navigerer multi-page workflows - Dynamisk tilpasning til UI-endringer - Cross-application (web og desktop) - Natural language interface -- **Regioner:** East US 2, Sweden Central, South India — **IKKE Norway East** -- Kontekstvindu: 8,192 tokens, maks output: 1,024 tokens -- Registrering: `https://aka.ms/oai/cuaaccess` +- **Region/residens:** verifiser deployment-region for `gpt-5.4` mot region-availability — den utgåtte `computer-use-preview`-modellen var begrenset til East US 2 / Sweden Central / South India (IKKE Norway East) +- Kontekst/output: følger `gpt-5.4` (de tidligere `computer-use-preview`-tallene 8 192 / 1 024 er utgått) +- Registrering (limited access, gpt-5.4): `https://aka.ms/OAI/gpt54access` ### 6. Deep Research tool **Status:** Deprecated — det dedikerte Deep Research-*toolet* er utfaset. Bruk i stedet `o3-deep-research`-**modellen** via web search-toolet eller en MCP-tool (Responses API, `2025-11-15-preview`). Selve modellen er fortsatt aktiv. @@ -289,7 +289,7 @@ Tilgjengelig i: - God dekning **via Global** — men Norge-residens er begrenset - Azure OpenAI tilgjengelig via Global Standard (data prosesseres globalt): GPT-4o, GPT-4.1-serien, o3, o4-mini, o3-mini, o1, GPT-5-familien - **Norge-resident (data forblir i Norge — Standard/Regional PTU):** KUN `gpt-4o` (Standard) + `gpt-4o`/`gpt-4o-mini` (Regional PTU). `gpt-4.1`, o-serien og hele GPT-5-familien er **IKKE** regionale i Norway East (verifisert 2026-06-18 mot Learn region-tabeller) -- GPT-5-familien: kun Global Standard/Provisioned (data globalt). `gpt-5.5` også via Data Zone Standard (EU-residens) — eneste GPT-5 med residens-tier i Norway East +- GPT-5-familien: kun Global Standard/Provisioned (data globalt). `gpt-5.4` og `gpt-5.5` via Data Zone Standard (EU-residens) — eneste GPT-5-modeller med residens-tier i Norway East - DeepSeek-R1, DeepSeek-V3-0324, DeepSeek-R1-0528 (Foundry Models) - Grok-4, Llama-modeller - Foundry Agent Service (GA) @@ -297,7 +297,7 @@ Tilgjengelig i: - Deep Research (`o3-deep-research`) — **Norway East er ett av kun TO regioner globalt** - Responses API (bekreftet) - **IKKE tilgjengelig i Norway East:** - - Computer-Use (`computer-use-preview`) — kun East US 2, Sweden Central, South India + - Computer Use: `computer-use-preview` er utgått (computer use bruker nå `gpt-5.4` + computer-tool) — verifiser gjeldende region; preview-modellen var kun East US 2 / Sweden Central / South India - Sora video generation — kun East US 2 og Sweden Central - GPT-image-1 — begrenset tilgang @@ -364,9 +364,9 @@ Microsoft.CognitiveServices/account (kind: AIServices) ### Norway East-spesifikke råd - Deep Research er **bedre egnet for Norway East** enn Sweden Central (ett av kun to regioner) -- Computer-Use krever deployment til Sweden Central eller East US 2 +- Computer Use: `computer-use-preview` utgått; bruker nå `gpt-5.4` + computer-tool — verifiser deployment-region (preview-modellen var Sweden Central / East US 2) - GPT-4.1 og DeepSeek-modeller er tilgjengelig, men **kun via Global** (ikke Norge-resident) -- For streng Norge-residens: KUN `gpt-4o`/`gpt-4o-mini` (Standard/Regional PTU) — `gpt-4.1`, o-serien og hele GPT-5-familien er kun Global i Norway East. For EU-residens med nyere modell: `gpt-5.5` via Data Zone Standard, eller flytt til Sweden Central (verifisert 2026-06-18) +- For streng Norge-residens: KUN `gpt-4o`/`gpt-4o-mini` (Standard/Regional PTU) — `gpt-4.1`, o-serien og hele GPT-5-familien er kun Global i Norway East. For EU-residens med nyere modell: `gpt-5.4`/`gpt-5.5` via Data Zone Standard, eller flytt til Sweden Central (verifisert 2026-06-29) ### Spørsmål å stille kunden - "Trenger dere å sammenligne ulike AI-modeller, eller er GPT tilstrekkelig?" diff --git a/skills/ms-ai-advisor/references/platforms/model-catalog-2026.md b/skills/ms-ai-advisor/references/platforms/model-catalog-2026.md index 74a93a8..aa70a75 100644 --- a/skills/ms-ai-advisor/references/platforms/model-catalog-2026.md +++ b/skills/ms-ai-advisor/references/platforms/model-catalog-2026.md @@ -23,7 +23,7 @@ Modellkatalogen er delt i to kategorier: OpenAIs flaggskip reasoning-modeller. Alle versjonene støtter Chat Completions API, Responses API, structured outputs, text/image input og parallel tool calling. **Tilgangsmodell:** -- `gpt-5`, `gpt-5-codex`, `gpt-5-pro` — krever registrering: `https://aka.ms/oai/gpt5access` +- `gpt-5`, `gpt-5-codex`, `gpt-5-pro` — åpen tilgang (tilgang ikke lenger begrenset; verifisert mot Microsoft Learn 2026-06) - `gpt-5-mini`, `gpt-5-nano`, `gpt-5-chat` — åpen tilgang, ingen registrering - `gpt-5.5` — enkelte quota-tiers krever quota-forespørsel (Tier 5/6 har quota by default) @@ -41,7 +41,7 @@ OpenAIs flaggskip reasoning-modeller. Alle versjonene støtter Chat Completions | `gpt-5.3-chat` | 2026-03-03 (Preview) | 128K (input 111K) | 16 384 | Conversation-variant | | `gpt-5.4` | 2026-03-05 | **1 050 000** | 128K | Kontekstsprang til ~1M; computer use. + `-mini`/`-nano` (2026-03-17) | | `gpt-5.4-pro` | 2026-03-05 | 1 050 000 | 128K | Høyeste kapabilitet (ikke lenger restricted) | -| `gpt-5.5` | 2026-04-24 | 1 050 000 (input 922K) | 128K | Treningsdata des. 2025; **eneste GPT-5 med Data Zone i Norway East** | +| `gpt-5.5` | 2026-04-24 | 1 050 000 (input 922K) | 128K | Treningsdata des. 2025; **Data Zone i Norway East (sammen med gpt-5.4)** | | `gpt-chat-latest` | 2026-05-28 (Preview) | 128K (input 111K) | 16 384 | «GPT-5.5 Instant» (OpenAI `chat-latest`); kun Global Standard | **PTU-ratio for GPT-5:** 1 output token teller som 8 input tokens mot utnyttelsesgrensen. @@ -192,7 +192,7 @@ Microsofts egne SLM-er (Small Language Models), optimalisert for effektiv infere | Modell | Kontekst (input) | Max output | PTU: input TPM/PTU | Latency target (99%) | Norway East — residens | Tilgang | |--------|-----------------|------------|---------------------|---------------------|------------------------|---------| -| `gpt-5` | 272K | 128K | 4 750 | >50 TPS | Kun Global | Registrering | +| `gpt-5` | 272K | 128K | 4 750 | >50 TPS | Kun Global | Åpen | | `gpt-5-mini` | 272K | 128K | 23 750 | >80 TPS | Kun Global | Åpen | | `gpt-5-nano` | 272K | 128K | Høy | >100 TPS | Kun Global Std | Åpen | | `gpt-5.5` | 922K | 128K | — | — | Data Zone (EU) + Global Prov | Quota-tier | @@ -261,7 +261,7 @@ Garantert kapasitet med forutsigbar latens. Egnet for: |--------|--------------------------------| | GPT-5-serien | 8 input tokens | | GPT-4.1-serien | 4 input tokens | -| DeepSeek-R1/V3 | 1 (standard, som input) | +| DeepSeek-R1/V3 | 4 input tokens | | Llama-3.3-70B | 4 input tokens (unntak fra standard) | ### Fine-tuning-kostnader @@ -297,7 +297,7 @@ Ved bruk av fine-tuned modeller: | `gpt-5.2` (+ `-codex`) | – | – | ✅ | ✅ | | `gpt-5.3-codex` | – | – | ✅ | ✅ | | `gpt-5.3-chat` | – | – | ✅ | – | -| `gpt-5.4` | – | – | ✅ | ✅ | +| `gpt-5.4` | – | ✅³ | ✅ | ✅ | | `gpt-5.4-mini` / `-nano` | – | – | ✅ | – | | `gpt-5.4-pro` | – | – | – | – | | `gpt-5.5` | – | ✅³ | – | ✅ | @@ -306,9 +306,9 @@ Ved bruk av fine-tuned modeller: ¹ Norge-resident = Standard/Regional **eller** Regional Provisioned (PTU) i Norway East — prosessering forblir i Norge. ² `o3-deep-research` finnes i Norway East kun via **Global Standard** (Norway East + West US er de eneste to regionene globalt) — data prosesseres globalt, ikke resident. -³ `gpt-5.5` er den **eneste** modellen med Data Zone Standard (EU-residens) i Norway East. +³ `gpt-5.4` og `gpt-5.5` er de eneste modellene med Data Zone Standard (EU-residens) i Norway East (verifisert 2026-06-29 mot Learn region-tabell, Data Zone standard-pivot, Europa). -**Konsekvens (viktig korreksjon 2026-06-18):** De **eneste generative modellene som kan kjøres Norge-resident** i Norway East er **`gpt-4o` og `gpt-4o-mini`**. `gpt-4.1`, hele o-serien (`o3`/`o4-mini`/`o3-mini`/`o1`) og hele GPT-5-familien har **ingen** Standard/Regional- eller Regional-PTU-deployment i Norway East — de kjører kun Global (data globalt), eller for `gpt-5.5` via Data Zone (EU). I tillegg er `text-embedding-3-large`, `text-embedding-ada-002` og `whisper` Norge-resident (Standard) i Norway East. +**Konsekvens (viktig korreksjon 2026-06-18):** De **eneste generative modellene som kan kjøres Norge-resident** i Norway East er **`gpt-4o` og `gpt-4o-mini`**. `gpt-4.1`, hele o-serien (`o3`/`o4-mini`/`o3-mini`/`o1`) og hele GPT-5-familien har **ingen** Standard/Regional- eller Regional-PTU-deployment i Norway East — de kjører kun Global (data globalt), eller for `gpt-5.4`/`gpt-5.5` via Data Zone (EU). I tillegg er `text-embedding-3-large`, `text-embedding-ada-002` og `whisper` Norge-resident (Standard) i Norway East. **Foundry Models sold directly by Azure (DeepSeek, Llama, Mistral, Grok):** tilgjengelig i Norway East kun via **Global Standard** (data prosesseres globalt) — ikke egnet for dataresidens-krav. Omfatter DeepSeek-R1/R1-0528/V3-0324/V3.1/V3.2, Llama-4-Maverick/Llama-3.3-70B, Grok-4, MAI-DS-R1, mistral-document-ai. @@ -329,7 +329,7 @@ Bred dekning via Global Standard/Provisioned (GPT-5-familien, GPT-4.1, o-serien, | `o4-mini` — resident | ❌ (kun Global) | ✅ | | `o3` — resident | ❌ (kun Global) | ❌ (kun Global) | | GPT-5-familien — resident | ❌ | delvis (`gpt-5.1` Standard) | -| `gpt-5.5` — Data Zone (EU-resident) | ✅ (eneste i Norden) | ❌ | +| `gpt-5.4`/`gpt-5.5` — Data Zone (EU-resident) | ✅ | ❌ | | `o3-deep-research` (Global) | ✅ | ❌ | | `computer-use-preview` / `sora-2` / `gpt-image-2` | ❌ | ✅ (Global) | | Dataresidens for resident-modeller | Norsk | Svensk | @@ -357,7 +357,7 @@ Har kunden Norway East-krav (dataresidens)? │ IKKE Norge-resident: gpt-4.1, o3/o4-mini/o3-mini/o1, HELE GPT-5-familien (kun Global her) │ Trenger du nyere modell enn gpt-4o? → løs residens i Sweden Central, eller aksepter EU-sone ├── Ja, men EU-sone-residens er akseptabelt (data innen EU, ikke garantert Norge) -│ → gpt-5.5 via Data Zone Standard i Norway East (eneste GPT-5 med Data Zone her) +│ → gpt-5.4/gpt-5.5 via Data Zone Standard i Norway East (eneste GPT-5-modeller med Data Zone her) │ eller Sweden Central Standard for gpt-4.1 / o4-mini / gpt-5.1 (Sverige-resident) │ DeepSeek/øvrige har ingen Data Zone i Norway East → Global Standard (data globalt) └── Nei (global prosessering OK) → hele GPT-5-familien (5/5.1/5.2/5.3/5.4) via Global Standard/Provisioned, diff --git a/skills/ms-ai-advisor/references/prompt-engineering/chain-of-thought-prompting.md b/skills/ms-ai-advisor/references/prompt-engineering/chain-of-thought-prompting.md index 8008853..fa8b6c5 100644 --- a/skills/ms-ai-advisor/references/prompt-engineering/chain-of-thought-prompting.md +++ b/skills/ms-ai-advisor/references/prompt-engineering/chain-of-thought-prompting.md @@ -65,7 +65,7 @@ Disse modellene utfører **innebygd resonnering** automatisk og returnerer både |----------------|-------------------------|---------------------|---------|--------------| | **Automatisk reasoning** | ✅ | ✅ | ✅ | ✅ | | **reasoning_effort parameter** | ✅ (low/medium/high) | ✅ (low/medium/high) | ✅ (low/medium/high) | ✅ (none/minimal/low/medium/high/xhigh) | -| **reasoning_summary** | ❌ | ✅ (limited access) | ✅ (limited access) | ✅ (auto/concise/detailed) | +| **reasoning_summary** | ❌ | ✅ (limited access) | ✅ (limited access) | ✅ (auto/detailed — GPT-5-serien støtter ikke `concise`) | | **Developer messages** | ✅ | ✅ | ✅ | ✅ | | **Streaming** | ❌ (o1, o1-preview) | ✅ (limited access for o3) | ✅ | ✅ | @@ -90,7 +90,7 @@ response = client.responses.create( model="gpt-5", # replace with model deployment name reasoning={ "effort": "medium", - "summary": "auto" # auto, concise, or detailed + "summary": "auto" # auto eller detailed (gpt-5-serien støtter ikke concise) }, text={ "verbosity": "low" # New with GPT-5 models diff --git a/skills/ms-ai-advisor/references/prompt-engineering/real-time-reasoning-performance.md b/skills/ms-ai-advisor/references/prompt-engineering/real-time-reasoning-performance.md index 97b4a31..ace21f3 100644 --- a/skills/ms-ai-advisor/references/prompt-engineering/real-time-reasoning-performance.md +++ b/skills/ms-ai-advisor/references/prompt-engineering/real-time-reasoning-performance.md @@ -110,7 +110,7 @@ for chunk in completion: - `gpt-realtime-mini` (2025-12-15) - `gpt-realtime-1.5` (2026-02-23) -**Deployment regions:** East US 2, Sweden Central (global deployments). +**Deployment:** Tilgjengelig for global deployments (ikke begrenset til East US 2 / Sweden Central; verifisert mot Microsoft Learn 2026-06). **API-endepunkt:** Bruk GA-endepunktet med `/openai/v1` i URL-en. Realtime API bruker **ikke** lenger date-baserte `api-version`-verdier; eldre preview-endepunkter (`/openai/realtimeapi/sessions`) er erstattet av GA-stiene `/openai/v1/realtime/...`. @@ -159,7 +159,7 @@ for chunk in completion: - Mindre sannsynlig å "chunke" transkripsjon før bruker er ferdig. - Bedre for speech-to-speech samtaler (venter på naturlig pause). -**Konfidensmarkering:** Middels (⚠️) — Realtime API er fortsatt i public preview (per januar 2026). Produksjonsbruk krever risikovurdering. +**Konfidensmarkering:** Høy (✅) — Realtime API bruker GA-endepunktet, og GA-modellene `gpt-realtime` / `gpt-realtime-1.5` er tilgjengelige (kun de eldre `gpt-4o-realtime-preview`-modellene er preview). Produksjonsbruk krever fortsatt risikovurdering. --- @@ -365,13 +365,13 @@ Deployment C: Chatbot (variabel prompt, medium output) ### Schrems II og Data Residency -**Realtime API regions (per januar 2026):** East US 2, Sweden Central. +**Realtime-modellenes deployment (verifisert mot Microsoft Learn 2026-06):** GPT realtime-modeller tilbys som **global deployments** («The GPT real-time models are available for global deployments»). De er altså ikke begrenset til East US 2 / Sweden Central, men global deployment innebærer at inferens kan rutes til Azure-kapasitet utenfor EU. **Konsekvens:** -- **Sweden Central:** EU-region, bedre for GDPR-compliance (men fortsatt USA-eid selskap). -- **East US 2:** USA-region, kan kreve DPIA for offentlig sektor. +- **Global deployment gir ikke EU-dataresidens i seg selv** — data kan behandles utenfor EU/EØS. For norsk offentlig sektor utløser dette en Schrems II-/overføringsvurdering (TIA) og som regel DPIA. +- Den tidligere antakelsen om å «pinne» realtime til Sweden Central for EU-residens holder ikke for en global deployment. Sjekk audio-modellenes deployment-type/region i [region-availability-kilden](https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure-region-availability) før en residens-konklusjon trekkes. -**Anbefaling:** Vurder Sweden Central for norsk offentlig sektor hvis Realtime API er kritisk. For standard completions, bruk Norway East (GPT-4o/GPT-4o mini tilgjengelig der). +**Anbefaling:** Avklar faktisk dataresidens for de spesifikke realtime-modellene mot region-availability-kilden FØR produksjonsbruk i offentlig sektor; behandle realtime som global deployment i risikovurderingen. For standard completions er regionale deployments (f.eks. Norway East, GPT-4o/GPT-4o mini) tilgjengelig. ### Accessibility (Universell Utforming) @@ -427,7 +427,7 @@ Deployment C: Chatbot (variabel prompt, medium output) **❌ Ikke anbefal hvis:** - Kun text-basert chatbot (bruk standard streaming i stedet). -- Klient har strenge data residency-krav og kan ikke bruke East US 2 / Sweden Central. +- Klient har strenge dataresidens-krav som global deployment ikke oppfyller (realtime tilbys som global deployment — se Schrems II-seksjonen). - Budsjett er begrenset (audio tokens er 10-40x dyrere enn text). ### Typiske Spørsmål fra Klienter @@ -443,9 +443,9 @@ Deployment C: Chatbot (variabel prompt, medium output) **Q: "Er Realtime API production-ready for offentlig sektor?"** **A (per januar 2026):** -- **Teknisk:** Public preview, ikke SLA. -- **Data residency:** Sweden Central er EU-region (bedre enn USA). -- **Anbefaling:** Pilot i ikke-kritiske tjenester først. Vent på GA for produksjonsbruk i kritiske systemer. +- **Teknisk:** GA-endepunkt og GA-modeller (`gpt-realtime`, `gpt-realtime-1.5`); kun de eldre `gpt-4o-realtime-preview`-modellene er preview. +- **Data residency:** Realtime tilbys som global deployment — avklar faktisk dataresidens mot region-availability-kilden før produksjon (se Schrems II-seksjonen). +- **Anbefaling:** Risikovurdering og DPIA før produksjonsbruk i offentlig sektor; pilot i ikke-kritiske tjenester først. **Q: "Hvordan måle om streaming faktisk hjelper?"** @@ -504,10 +504,10 @@ Deployment C: Chatbot (variabel prompt, medium output) **Verification steps:** 1. ✅ **Streaming impact:** Bekreftet at `stream: true` reduserer TTFT men ikke total tid (dokumentasjon + code samples). -2. ✅ **Realtime API models:** Bekreftet at `gpt-4o-mini-realtime-preview` og `gpt-4o-realtime-preview` er tilgjengelige i East US 2 / Sweden Central. +2. ✅ **Realtime API models:** Bekreftet at GPT realtime-modeller tilbys for global deployments (ikke begrenset til East US 2 / Sweden Central); GA-modeller `gpt-realtime` / `gpt-realtime-1.5` i tillegg til eldre `gpt-4o-realtime-preview`. 3. ✅ **VAD modes:** Bekreftet at `server_vad`, `semantic_vad`, og `none` er supported turn detection types. 4. ✅ **Latency metrics:** Bekreftet at Time to Response (TTFT) og Average Token Generation Rate er recommended metrics for streaming. 5. ✅ **Speech latency:** first byte client latency og AudioDataStream-streaming bekreftet. Text streaming via WebSocket v2 bekreftet for C#, Python. 5. ⚠️ **Pricing:** Audio token pricing ikke eksplisitt i dokumentasjon per januar 2026. Brukt representative estimates basert på historisk OpenAI pricing structure. -**Confidence level:** Høy (✅) for tekniske detaljer, Middels (⚠️) for pricing og production-readiness av Realtime API (public preview). +**Confidence level:** Høy (✅) for tekniske detaljer, Middels (⚠️) for pricing. Realtime API bruker GA-endepunkt/GA-modeller (ikke lenger public preview). diff --git a/skills/ms-ai-advisor/references/prompt-engineering/reasoning-models-o1-o3-optimization.md b/skills/ms-ai-advisor/references/prompt-engineering/reasoning-models-o1-o3-optimization.md index 7e43131..8cc15ba 100644 --- a/skills/ms-ai-advisor/references/prompt-engineering/reasoning-models-o1-o3-optimization.md +++ b/skills/ms-ai-advisor/references/prompt-engineering/reasoning-models-o1-o3-optimization.md @@ -423,8 +423,7 @@ Reasoning-tokens faktureres til **output-raten** — de er ikke en egen prisklas Reasoning models krever: - **Azure OpenAI Service** subscription (ingen spesielle lisenser) -- **Limited access request** for enkelte modeller (o3-pro, gpt-5-pro, gpt-5-codex) -- Request via: https://aka.ms/oai/o3access eller https://aka.ms/oai/gpt5access +- Ingen «limited access request» lenger — tilgang til o3-pro, gpt-5-pro og gpt-5-codex er ikke lenger begrenset (verifisert mot Microsoft Learn 2026-06) **Ingen spesielle lisenskrav:** - `o1`, `o3-mini`, `o4-mini`, `codex-mini` diff --git a/skills/ms-ai-engineering/references/azure-ai-services/ai-services-cost-optimization.md b/skills/ms-ai-engineering/references/azure-ai-services/ai-services-cost-optimization.md index 3b2ff01..6f060c4 100644 --- a/skills/ms-ai-engineering/references/azure-ai-services/ai-services-cost-optimization.md +++ b/skills/ms-ai-engineering/references/azure-ai-services/ai-services-cost-optimization.md @@ -62,7 +62,7 @@ PTU er en kapasitetsbasert prismodell for Azure OpenAI, primært for produksjons **Fakturering:** - **Timepris:** Beregnes per PTU per time ($/PTU/hr) - **Pro-rata:** Delvis time faktureres proporsjonalt (15 min = 1/4 timepris) -- **Reservasjonsrabatt:** 1-års eller 3-års Azure Reservations gir betydelige rabatter (opptil 50 % besparelse) +- **Reservasjonsrabatt:** 1-måneds eller 1-års Azure Reservations gir betydelige rabatter (ingen 3-årstermin for Foundry PTU; faktisk rabatt varierer per modellfamilie og term) **Kapasitetsplanlegging:** - Bruk **Foundry PTU Calculator** (tilgjengelig i Microsoft Foundry portal) @@ -109,7 +109,7 @@ PTU er en kapasitetsbasert prismodell for Azure OpenAI, primært for produksjons - Vurder PTU for Azure OpenAI med SLA-krav **Fase 3 (Optimalisering):** Reservasjoner + Tagging -- Kjøp 1-års eller 3-års PTU-reservasjon +- Kjøp 1-måneds eller 1-års PTU-reservasjon - Bruk tags for kostnadsallokering per prosjekt/team **Verified** – Microsoft Learn: Plan and Manage Costs for Azure OpenAI. @@ -156,7 +156,7 @@ PTU er en kapasitetsbasert prismodell for Azure OpenAI, primært for produksjons - Azure OpenAI i produksjon - SLA-krav (latency, throughput) - Høy, forutsigbar trafikk (> 100K tokens/dag) -- Langsiktig forpliktelse (1-3 år reservasjon gir best ROI) +- Langsiktig forpliktelse (1-års reservasjon gir best ROI; 3-årstermin finnes ikke for Foundry PTU) ❌ **Ikke bruk når:** - Lav trafikk eller pilot-fase @@ -307,7 +307,7 @@ PTU er en kapasitetsbasert prismodell for Azure OpenAI, primært for produksjons 1. "Er dette Azure OpenAI i produksjon?" 2. "Har dere latency/throughput-krav i SLA?" 3. "Er trafikken forutsigbar (> 100K tokens/dag)?" -4. "Kan dere forplikte deg til 1-3 år?" +4. "Kan dere forplikte deg til 1 år?" **Anbefaling:** - Hvis JA på alle → **Anbefal PTU med 1-års reservasjon** diff --git a/skills/ms-ai-engineering/references/azure-ai-services/ai-services-vs-foundry-tools-selection.md b/skills/ms-ai-engineering/references/azure-ai-services/ai-services-vs-foundry-tools-selection.md index 2bdf3c6..e2ce199 100644 --- a/skills/ms-ai-engineering/references/azure-ai-services/ai-services-vs-foundry-tools-selection.md +++ b/skills/ms-ai-engineering/references/azure-ai-services/ai-services-vs-foundry-tools-selection.md @@ -106,7 +106,7 @@ Microsoft tilbyr flere nivåer av AI-tjenester under paraplynavnet "Azure AI Ser **Nøkkelkapabiliteter:** - **Agent Service:** Managed runtime for agentic AI (conversation state, tool orchestration, safety enforcement) -- **Model Catalog:** 100+ modeller fra Microsoft, OpenAI, Anthropic, Meta, Mistral, Cohere +- **Model Catalog:** over 1 900 modeller fra Microsoft, OpenAI, Anthropic, Meta, Mistral, Cohere - **Connected agents:** Integrasjon med Azure AI Search, SharePoint, Bing, Azure Functions, Logic Apps - **Workflows:** YAML-basert multi-agent orkestrering med visual designer - **Observability:** Built-in tracing via Application Insights (traces, evaluations, conversation-level visibility) @@ -124,17 +124,16 @@ Microsoft tilbyr flere nivåer av AI-tjenester under paraplynavnet "Azure AI Ser **Definisjon:** Spesialisert tjeneste for å få tilgang til OpenAI-modeller (GPT, DALL-E, Whisper, Embeddings) med Azure enterprise-fordeler (sikkerhet, compliance, SLA). -#### Modellserie (2026-02) +#### Modellserie (2026-06) | Modell | Bruksområde | Deployment-typer | |--------|-------------|------------------| -| **o4-mini** | Reasoning, kompleks problemløsning | Standard, Global Standard | -| **o3, o3-mini** | Avansert reasoning | Standard, Provisioned Throughput | +| **GPT-5-generasjonen** (GPT-5/5.1/5.2/5.4/5.5 + -mini/-nano/-codex) | Ledende reasoning og chat (flaggskip; GPT-5.5 nyest) | Se region-verifisert `model-catalog-2026.md` | +| **GPT-4.1, GPT-4.1-mini, GPT-4.1-nano** | Lang kontekst (~1M tokens), kostnadsspekter | Standard, Global Standard, Provisioned | +| **o4-mini, o3, o3-mini** | Eldre o-serie reasoning (erstattet av GPT-5-gen for nye design) | Standard, Provisioned | | **GPT-4o, GPT-4o-mini** | Chat, multimodal (tekst/bilde) | Standard, Global Standard, Provisioned | -| **GPT-4 Turbo** | Long-context tasks (128k tokens) | Standard, Provisioned | -| **GPT-3.5-Turbo** | Kostnadseffektiv chat | Standard, Global Standard | -| **DALL-E 3** | Bildegenerering | Standard | -| **Whisper** | Speech-to-text | Standard | +| **Bildegenerering (GPT-image / DALL-E 3)** | Bildegenerering | Standard | +| **Whisper / gpt-4o-transcribe** | Tale-til-tekst | Standard | | **Embeddings** (text-embedding-3) | Vektorisering for RAG | Standard | #### Deployment-typer @@ -529,7 +528,7 @@ START: Hvilken AI-kapabilitet trenger du? **Faktorer som påvirker kostnad:** 1. **Volum:** Antall API-kall, tokens, bilder, minutter audio -2. **Modellvalg:** GPT-4o > GPT-4o-mini > GPT-3.5-Turbo (kostnad) +2. **Modellvalg:** GPT-4o > GPT-4o-mini > GPT-4.1-nano (kostnad) 3. **Deployment-type:** PTU > Standard > Global Standard 4. **Region:** Noen regioner er dyrere (f.eks. EU vs. US East) 5. **Features:** Content Safety, fine-tuning, hosting (Azure OpenAI) diff --git a/skills/ms-ai-engineering/references/mlops-genaiops/genaiops-llm-specific-practices.md b/skills/ms-ai-engineering/references/mlops-genaiops/genaiops-llm-specific-practices.md index 4f00492..9592a63 100644 --- a/skills/ms-ai-engineering/references/mlops-genaiops/genaiops-llm-specific-practices.md +++ b/skills/ms-ai-engineering/references/mlops-genaiops/genaiops-llm-specific-practices.md @@ -143,7 +143,7 @@ MLflow Tracing provides end-to-end observability for GenAI applications: **Hva:** Unified platform for GenAI lifecycle management. **GenAIOps capabilities:** -- **Model Catalog**: Browse 1600+ foundation models (OpenAI, Meta, Mistral, Cohere) +- **Model Catalog**: Browse over 1,900 models (OpenAI, Meta, Mistral, Cohere) - **Prompt Flow**: Visual designer for LLM workflows - **Evaluation SDK**: Built-in evaluators (groundedness, relevance, coherence, fluency, safety) - **Content Safety**: Real-time filtering (hate, violence, sexual, self-harm) diff --git a/skills/ms-ai-engineering/references/mlops-genaiops/llm-evaluation-production.md b/skills/ms-ai-engineering/references/mlops-genaiops/llm-evaluation-production.md index 08c03c4..67f7eba 100644 --- a/skills/ms-ai-engineering/references/mlops-genaiops/llm-evaluation-production.md +++ b/skills/ms-ai-engineering/references/mlops-genaiops/llm-evaluation-production.md @@ -579,9 +579,6 @@ MLflow 3 (SDK `mlflow[databricks]>=3.1`) introduces a unified evaluation model: | `RetrievalGroundedness` | No | Hallucination detection | | `Safety` | No | Harmful/toxic content | | `Correctness` | Yes | Accuracy vs ground truth | -| `Completeness` | Yes | All questions addressed | -| `Fluency` | No | Grammatically correct and naturally flowing | -| `Equivalence` | Yes | Response equivalent to expected output | | `RetrievalSufficiency` | Yes | Context provides all necessary information | | `ToolCallCorrectness` | Yes | Tool calls and arguments | | `ToolCallEfficiency` | No | Redundant tool usage | @@ -1112,7 +1109,7 @@ Dette området utvikler seg raskt. Anbefalt re-verification: - **Annually:** Compliance requirements (AI Act implementation guidance evolves) **Siste research-dato:** 2026-06-19 -**Kilder brukt:** 7 Microsoft Learn articles, 15 code samples, Azure AI Evaluation SDK v1.14.0 +**Kilder brukt:** 7 Microsoft Learn articles, 15 code samples, Azure AI Evaluation SDK v1.17.0 --- diff --git a/skills/ms-ai-engineering/references/rag-architecture/embedding-models-selection.md b/skills/ms-ai-engineering/references/rag-architecture/embedding-models-selection.md index ffcc7fd..d72dcc8 100644 --- a/skills/ms-ai-engineering/references/rag-architecture/embedding-models-selection.md +++ b/skills/ms-ai-engineering/references/rag-architecture/embedding-models-selection.md @@ -361,7 +361,7 @@ for doc in search_result: ### Lisensiering - **Azure OpenAI Service:** Krever Azure-abonnement, ingen separate lisenser for embedding-modeller -- **Azure AI Search:** Basic tier fra $75/mnd (1 GB storage), Standard S1 fra $250/mnd (25 GB storage) +- **Azure AI Search:** Basic tier fra $75/mnd (15 GB storage per partisjon for nye services; eldre: 2 GB), Standard S1 fra $250/mnd (160 GB; eldre: 25 GB) - **Custom embeddings (self-hosted):** Ingen lisenskostnad utover compute (Azure ML, Kubernetes) --- diff --git a/skills/ms-ai-engineering/references/rag-architecture/rag-context-windows.md b/skills/ms-ai-engineering/references/rag-architecture/rag-context-windows.md index 56e2161..95d5ca4 100644 --- a/skills/ms-ai-engineering/references/rag-architecture/rag-context-windows.md +++ b/skills/ms-ai-engineering/references/rag-architecture/rag-context-windows.md @@ -10,7 +10,7 @@ Context window-størrelse er en av de mest kritiske faktorene som bestemmer kvaliteten og kostnaden for en RAG-løsning. En language model har en begrenset kapasitet for tokens den kan prosessere i en enkelt request — dette omtales som modellens context window. For RAG-implementasjoner må man balansere mellom å gi modellen nok kontekst til å generere gode svar, uten å overbelaste context window eller sløse med tokens (som koster penger). -Med innføringen av long-context models som GPT-4 Turbo (128k tokens) og GPT-4.1 (context windows opp til 200k+ tokens), har arkitekter fått nye muligheter: skal man fortsatt bruke klassisk RAG med retrieval av små, relevante chunks, eller kan man nå sende hele dokumenter direkte til modellen? Svaret avhenger av use case, kostnad, latency-krav og modellens faktiske evne til å utnytte store context windows — kjent som "lost-in-the-middle"-problemet. +Med innføringen av long-context models som GPT-4 Turbo (128k tokens) og GPT-4.1 (context windows opp til ~1M tokens), har arkitekter fått nye muligheter: skal man fortsatt bruke klassisk RAG med retrieval av små, relevante chunks, eller kan man nå sende hele dokumenter direkte til modellen? Svaret avhenger av use case, kostnad, latency-krav og modellens faktiske evne til å utnytte store context windows — kjent som "lost-in-the-middle"-problemet. I denne kunnskapsreferansen dekkes token budgeting, context window management, og når man skal velge RAG-basert chunking versus long-context direct prompting, med fokus på Azure OpenAI-modeller og integrasjon i Microsoft-stakken. @@ -29,9 +29,9 @@ I denne kunnskapsreferansen dekkes token budgeting, context window management, o **Verified (Azure OpenAI):** - GPT-4 Turbo: 128k tokens context window -- GPT-4.1 series: Opp til 200k+ tokens context window +- GPT-4.1 series: ~1M tokens context window (maks 1 047 576; standard-deployments 300k, provisioned managed/batch 128k) - GPT-4o: 128k tokens -- o1-series: Varierende (o1: 200k, o3-mini: 128k) +- o1-series: Varierende (o1: 200k, o3-mini: 200k input / 100k output) ### Token Budget Allocation @@ -49,7 +49,7 @@ For GPT-4 Turbo (128k context): - Retrieved context: Max 30-50k tokens (ikke hele window) - Response buffer: 2-4k tokens -For GPT-4.1 (200k context): +For GPT-4.1 (~1M context): - Retrieved context: Kan økes til 100k tokens, men kvalitet avtar (lost-in-the-middle) - Reserved: 10-15k tokens @@ -82,7 +82,7 @@ Studier viser at LLMs har svakere performance når relevant informasjon er plass | Modell | Context Window | Anbefalt RAG-strategi | |--------|---------------|----------------------| | GPT-4 Turbo | 128k | Klassisk RAG (top-10 til top-50 chunks) | -| GPT-4.1 | 200k+ | Hybrid: RAG for precision queries, long-context for exploratory | +| GPT-4.1 | ~1M | Hybrid: RAG for precision queries, long-context for exploratory | | GPT-4o | 128k | Klassisk RAG med hybrid search | | o1-series | 200k | Long-context for reasoning tasks, RAG for factual grounding | @@ -200,14 +200,16 @@ Sjeldnere brukt, men effektiv for spesialiserte domener. ### Azure OpenAI Context Limits (Verified) -| Modell | Context Window | TPM (Default tier) | TPM (Enterprise tier) | -|--------|---------------|-------------------|----------------------| -| gpt-4 (turbo-2024-04-09) | 128k | 450k | 2M | -| gpt-4.1 | 200k+ | 1M | 5M | -| gpt-4o | 128k | 450k | 30M | -| gpt-4o-mini | 128k | 2M | 150M | -| o1 | 200k | 3M | 30M | -| o3-mini | 128k | 5M | 50M | +| Modell | Context Window | +|--------|---------------| +| gpt-4 (turbo-2024-04-09) | 128k | +| gpt-4.1 | ~1M (1 047 576) | +| gpt-4o | 128k | +| gpt-4o-mini | 128k | +| o1 | 200k | +| o3-mini | 200k input / 100k output | + +**TPM-grenser (Quota Tiers):** Det tidligere to-nivå-systemet «Default tier / Enterprise tier» er erstattet av **Quota Tiers** — Free Tier (Tier 0) pluss Tier 1–6 (Tier 6 høyest), med kvote som øker automatisk med forbruk. TPM/RPM defineres per region, per subscription og per modell/deployment-type. Eksempel GlobalStandard TPM (Tier 1 → Tier 6): gpt-4.1 1M → 45M; gpt-4o-mini 2M → 225M; o1 3M → 48M; o3-mini 5M → 80M. Maks TPM (Tier 6) = 225M. Se [quotas-limits](https://learn.microsoft.com/azure/foundry/openai/quotas-limits) for full per-tier-tabell. **Viktig:** TPM (Tokens Per Minute) = Max tokens som kan prosesseres per minutt på deployment-nivå. Hvis du sender én request med 50k input tokens + 2k output tokens = 52k tokens → teller mot TPM. @@ -319,9 +321,9 @@ Hvis du har 10,000 queries per måned: **11,300 NOK/måned** (kun LLM-kostnad, i 3. **Batch processing** for ikke-interaktive workloads (Azure OpenAI Batch API: 50% rabatt) 4. **Monitorering**: Bruk Azure Monitor for å spore token usage per deployment -**Verified (Azure OpenAI Batch Quota):** -- gpt-4.1: 500M tokens per month (Enterprise tier), 30M (Default tier) -- gpt-4o: 500M (Enterprise), 30M (Default) +**Batch-kvote (Azure OpenAI — nå per Quota Tier; tidligere «Default/Enterprise»):** +- gpt-4.1: ~500M tokens/måned (øvre tier), ~30M (lavere tier) — ikke re-verifisert mot gjeldende per-tier-tabell, se quotas-limits +- gpt-4o: ~500M (øvre), ~30M (lavere) — samme forbehold --- diff --git a/skills/ms-ai-engineering/references/rag-architecture/rag-cost-optimization.md b/skills/ms-ai-engineering/references/rag-architecture/rag-cost-optimization.md index 9aa2ab6..b50ff33 100644 --- a/skills/ms-ai-engineering/references/rag-architecture/rag-cost-optimization.md +++ b/skills/ms-ai-engineering/references/rag-architecture/rag-cost-optimization.md @@ -24,12 +24,12 @@ Valg av Azure AI Search pricing tier er avgjørende for total kostnad: |------|----------|---------|-----------|-------------------| | **Free** | POC, testing | 50 MB | Begrenset | NOK 0 | | **Basic** | Små produksjonsløsninger | 15 GB (services opprettet etter april 2024; eldre: 2 GB) | Moderat | ~NOK 700 | Verified (MCP 2026-04) | -| **S1** | Standard produksjon | 25 GB/partition | Høy | ~NOK 2,500 | -| **S2** | Store løsninger | 100 GB/partition | Meget høy | ~NOK 10,000 | -| **S3 HD** | Multitenant, mange små indekser | 200 GB | Høy | ~NOK 20,000 | -| **L1/L2** | Storage-optimized, sjeldne queries | 1 TB+ | Lavere | ~NOK 15,000+ | +| **S1** | Standard produksjon | 160 GB/partition | Høy | ~NOK 2,500 | +| **S2** | Store løsninger | 512 GB/partition | Meget høy | ~NOK 10,000 | +| **S3 HD** | Multitenant, mange små indekser | 1 024 GB/partition | Høy | ~NOK 20,000 | +| **L1/L2** | Storage-optimized, sjeldne queries | 2 TB (L1) / 4 TB (L2) | Lavere | ~NOK 15,000+ | -**Viktig:** Services opprettet etter april 2024 får større partitions til samme pris. Basic-tier: 15 GB per partisjon (eldre services: 2 GB). S1: 25 GB per partisjon. Tier switching er nå støttet — du kan bytte mellom Basic og Standard (S1/S2/S3) direkte (config må ikke overstige target-tier; regionen må ha kapasitet). Verified (MCP 2026-04). +**Viktig:** Services opprettet etter april 2024 får større partitions til samme pris. Basic-tier: 15 GB per partisjon (eldre services: 2 GB). S1: 160 GB per partisjon. Tier switching er nå støttet — du kan bytte mellom Basic og Standard (S1/S2/S3) direkte (config må ikke overstige target-tier; regionen må ha kapasitet). Verified (MCP 2026-04). **NY prismodell — Serverless (Preview, MCP 2026-06):** Azure AI Search har nå to prismodeller — **Dedicated** (tabellen over, fast pris per Search Unit) og **Serverless (Preview)**, forbruksbasert (Compute Units/time + per-GB/mnd lagring), ingen compute-kost ved idle. Serverless passer sporadiske/bursty workloads og multitenant; Dedicated passer jevn, forutsigbar last. Per juni 2026 er Serverless Developer i preview (kun West Central US, Switzerland North, Japan East), uten SLA, og støtter ikke migrering til/fra Dedicated. **SU og CU er ikke utvekslbare** — ikke bruk SU-baserte kalkulatorer for Serverless. @@ -350,12 +350,12 @@ Metrics: |------|-------------------|-------------------|-------------------|------| | Free | 0.00 | 0 | 1 | 50 MB, 1 index limit | | Basic | ~1.00 | ~730 | 1-3 | 15 GB per partition (etter april 2024) | Verified (MCP 2026-04) | -| S1 | ~3.50 | ~2,555 | 1-36 | 25 GB per partition | -| S2 | ~13.50 | ~9,855 | 1-36 | 100 GB per partition | -| S3 | ~27.00 | ~19,710 | 1-36 | 200 GB per partition | +| S1 | ~3.50 | ~2,555 | 1-36 | 160 GB per partition | +| S2 | ~13.50 | ~9,855 | 1-36 | 512 GB per partition | +| S3 | ~27.00 | ~19,710 | 1-36 | 1 024 GB per partition | | S3 HD | ~27.00 | ~19,710 | 1-36 | Optimalisert for mange indekser | -| L1 | ~20.00 | ~14,600 | 1-12 | 1 TB per partition | -| L2 | ~40.00 | ~29,200 | 1-12 | 2 TB per partition | +| L1 | ~20.00 | ~14,600 | 1-36 | 2 TB per partition | +| L2 | ~40.00 | ~29,200 | 1-36 | 4 TB per partition | **Viktig:** Kostnader = Base rate × (replicas × partitions). Eks: S1 med 2 replicas og 2 partitions = 4 SU = NOK 10,220/mnd. diff --git a/skills/ms-ai-governance/references/monitoring-observability/endpoint-health-and-capacity-planning.md b/skills/ms-ai-governance/references/monitoring-observability/endpoint-health-and-capacity-planning.md index a04b6fd..b4957b4 100644 --- a/skills/ms-ai-governance/references/monitoring-observability/endpoint-health-and-capacity-planning.md +++ b/skills/ms-ai-governance/references/monitoring-observability/endpoint-health-and-capacity-planning.md @@ -22,10 +22,10 @@ Utfordringen for arkitekter er å balansere kostnad (monitoreringsdata lagres i |---------|-------------|-------------------|-----------| | `AzureOpenAIRequests` | Totalt antall API-kall over tid | PT1M (1 minutt) | Ja | | `AzureOpenAIAvailabilityRate` | `(Total Calls - Server Errors) / Total Calls` (%) | PT1M | Nei | -| `TokensGenerated` | Completion tokens generert | PT1M | Ja | +| `GeneratedTokens` | Completion tokens generert | PT1M | Ja | | `ActiveTokens` | Totale tokens (prompt + completion) | PT1M | Ja | -| `PTUUtilization` | Prosentvis bruk av PTU-kapasitet | PT1M | Ja | -| `ProcessingTime` | End-to-end latency (ms) | PT1M | Ja | +| `AzureOpenAIProvisionedManagedUtilizationV2` | Prosentvis bruk av PTU-kapasitet | PT1M | Nei | +| `AzureOpenAITimeToResponse` | Time to response (ms) | PT1M | Ja | **Viktig:** Platform metrics samles automatisk uten konfigurasjon, men for å analysere i Log Analytics må diagnostic settings aktiveres. @@ -38,14 +38,14 @@ Utfordringen for arkitekter er å balansere kostnad (monitoreringsdata lagres i | **Quota** | Regionbasert grense per modell/subscription | TPM | Forespørres via support | | **429 Throttling** | HTTP-responskode når rate limit overstiges | - | Implementer retry-logic | -**RPM/TPM-ratio varierer per modell:** +**RPM/TPM-ratio varierer per modell.** Kvote tildeles nå via **Quota Tiers** (Free Tier (Tier 0) + Tier 1–6), ikke «1 Unit Capacity». Forholdet mellom RPM og TPM er likevel modellspesifikt (verifisert mot Tier 1, GlobalStandard): -| Modell | 1 Unit Capacity | RPM | TPM | -|--------|----------------|-----|-----| -| Eldre chat-modeller | 1 | 6 | 1,000 | -| o1, o1-preview | 1 | 1 | 6,000 | -| o3 | 1 | 1 | 1,000 | -| o3-mini, o1-mini | 1 | 1 | 10,000 | +| Modell | RPM | TPM | +|--------|-----|-----| +| Eldre chat-modeller | 6 | 1,000 | +| o1, o1-preview | 1 | 6,000 | +| o3 | 1 | 1,000 | +| o3-mini, o1-mini | 1 | 10,000 | **Viktig:** Deployment TPM kan IKKE overskride subscription quota for den modellen i den regionen. @@ -172,7 +172,7 @@ PUT https://management.azure.com/subscriptions/{subscriptionId}/resourceGroups/{ **Mønster:** - Bruk PTU i stedet for Standard (pay-as-you-go) - PTU gir dedikert kapasitet med forutsigbar latens -- Overvåk `PTUUtilization`-metrikk for kapasitetsplanlegging +- Overvåk `AzureOpenAIProvisionedManagedUtilizationV2`-metrikk for kapasitetsplanlegging - Sett alert hvis utilization > 80% (signal om behov for oppgradering) **Fordeler:** @@ -189,7 +189,7 @@ PUT https://management.azure.com/subscriptions/{subscriptionId}/resourceGroups/{ ```kql AzureMetrics -| where MetricName == "PTUUtilization" +| where MetricName == "AzureOpenAIProvisionedManagedUtilizationV2" | where TimeGenerated > ago(5m) | summarize AvgUtilization = avg(Average) by Resource | where AvgUtilization > 80 diff --git a/skills/ms-ai-governance/references/monitoring-observability/token-usage-tracking-attribution.md b/skills/ms-ai-governance/references/monitoring-observability/token-usage-tracking-attribution.md index a567a18..944f961 100644 --- a/skills/ms-ai-governance/references/monitoring-observability/token-usage-tracking-attribution.md +++ b/skills/ms-ai-governance/references/monitoring-observability/token-usage-tracking-attribution.md @@ -57,8 +57,8 @@ Azure OpenAI samler automatisk token-baserte metrics: **Tilgjengelige metrics:** - `TokenTransaction` — Total token count (input + output) -- `PromptTokens` — Input tokens -- `CompletionTokens` — Output tokens +- `InputTokens` — Input tokens +- `OutputTokens` — Output tokens - `ProcessedPromptTokens` — Tokens faktisk prosessert (kan avvike ved caching) **Aksess via:** @@ -442,7 +442,7 @@ AzureMetrics | **PTU** | Track utilization percentage against reserved capacity | **PTU Metrics:** -- `PTUUtilization` — Percentage of reserved capacity used +- `AzureOpenAIProvisionedManagedUtilizationV2` — Percentage of reserved capacity used - `ProcessedPromptTokens` — Input tokens processed - Input TPM per PTU — Model-specific (f.eks. 8450 TPM for Llama-3.3-70B) @@ -455,7 +455,7 @@ AzureMetrics ```kusto AzureMetrics | where ResourceProvider == "MICROSOFT.COGNITIVESERVICES" -| where MetricName == "PTUUtilization" +| where MetricName == "AzureOpenAIProvisionedManagedUtilizationV2" | summarize AvgUtilization = avg(Average), MaxUtilization = max(Maximum) by Resource, bin(TimeGenerated, 1h) | extend EfficiencyStatus = case( diff --git a/skills/ms-ai-governance/references/responsible-ai/responsible-ai-training-awareness.md b/skills/ms-ai-governance/references/responsible-ai/responsible-ai-training-awareness.md index 4f6ccb9..8fa52c8 100644 --- a/skills/ms-ai-governance/references/responsible-ai/responsible-ai-training-awareness.md +++ b/skills/ms-ai-governance/references/responsible-ai/responsible-ai-training-awareness.md @@ -60,7 +60,7 @@ AI-landskap endrer seg raskt. En gang-opplæring er utilstrekkelig. Mekanismer i - **Månedlige "AI Ethics Drop-ins"** — case reviews av reelle AI-hendelser (både interne og eksterne) - **Role-based refreshers** — kvartalsvis oppdatering når nye Microsoft AI-features lanseres (f.eks. Copilot extensibility, nye modeller) - **Incident-driven learning** — når AI-systemer feiler eller produserer uønskede outputs, konverteres dette til læringscases -- **Certification renewal** — AI-sertifiseringer (AI-900, AI-102) har ikke formell utløpsdato, men organisasjoner bør kreve re-cert hvert 18-24 måned +- **Certification renewal** — AI-900 (Fundamentals) har ingen formell utløpsdato, mens AI-102 (Azure AI Engineer Associate) har 12-måneders fornyelsesfrekvens og pensjoneres 30. juni 2026 (kan ikke oppnås eller fornyes etter denne datoen). Organisasjoner bør spore fornyelseskrav per sertifisering, ikke anta én felles regel **Baseline:** Microsoft anbefaler at minimum 80 % av alle som arbeider med AI-systemer (design, utvikling, godkjenning) skal ha gjennomført strukturert Responsible AI-opplæring. diff --git a/skills/ms-ai-governance/references/responsible-ai/stakeholder-communication-ai-decisions.md b/skills/ms-ai-governance/references/responsible-ai/stakeholder-communication-ai-decisions.md index 2907c66..ebe957b 100644 --- a/skills/ms-ai-governance/references/responsible-ai/stakeholder-communication-ai-decisions.md +++ b/skills/ms-ai-governance/references/responsible-ai/stakeholder-communication-ai-decisions.md @@ -598,18 +598,18 @@ Norske offentlige etater må følge **Lov om offentlige anskaffelser**, **GDPR** ### Copilot Studio — Agent Observability **Lisensiering**: -- Copilot Studio: Fra $200 per tenant/måned (inkluderer 25 000 messages) -- Ekstra messages: $0.015 per message +- Copilot Studio: Fra $200 per tenant/måned (inkluderer 25 000 Copilot Credits) +- Ekstra forbruk måles i Copilot Credits (per-credit-sats ikke re-verifisert mot gjeldende billing-rates — tidligere oppgitt ~$0.015) - Agent observability (Microsoft Agent 365): Inkludert i Copilot Studio-lisens **Kostnadskomponenter**: - **Base subscription**: Ca. NOK 2 200/måned -- **Overage**: Ca. NOK 0.16 per ekstra message +- **Overage**: Ca. NOK 0.16 per ekstra Copilot Credit (estimat, ikke re-verifisert) - **Azure Log Analytics** (for centralized logging): Fra NOK 200/måned -**Estimat (eksempel — 50 000 messages/måned)**: +**Estimat (eksempel — 50 000 Copilot Credits/måned)**: - Base: NOK 2 200 -- Overage (25 000 messages): NOK 4 000 +- Overage (25 000 Copilot Credits): NOK 4 000 - Log Analytics: NOK 500 - **Total**: ~NOK 6 700/måned diff --git a/skills/ms-ai-infrastructure/references/bcdr/ai-foundry-disaster-recovery-planning.md b/skills/ms-ai-infrastructure/references/bcdr/ai-foundry-disaster-recovery-planning.md index e8da941..3a60a3a 100644 --- a/skills/ms-ai-infrastructure/references/bcdr/ai-foundry-disaster-recovery-planning.md +++ b/skills/ms-ai-infrastructure/references/bcdr/ai-foundry-disaster-recovery-planning.md @@ -8,7 +8,7 @@ ## Introduksjon -Microsoft Foundry (tidligere Azure AI Studio / Azure Machine Learning) er Microsofts sentrale plattform for utvikling, evaluering og deployering av AI-modeller og agenter. Plattformen tilbyr imidlertid ikke automatisk failover eller disaster recovery ut av boksen -- dette er eksplisitt dokumentert av Microsoft. Det betyr at organisasjoner i norsk offentlig sektor som bygger forretningskritiske AI-loesninger pa AI Foundry, ma planlegge og implementere sin egen DR-strategi. +Microsoft Foundry (tidligere Azure AI Studio / Azure AI Foundry) er Microsofts sentrale plattform for utvikling, evaluering og deployering av AI-modeller og agenter. Plattformen tilbyr imidlertid ikke automatisk failover eller disaster recovery ut av boksen -- dette er eksplisitt dokumentert av Microsoft. Det betyr at organisasjoner i norsk offentlig sektor som bygger forretningskritiske AI-loesninger pa AI Foundry, ma planlegge og implementere sin egen DR-strategi. Disaster recovery for AI Foundry-prosjekter er mer kompleks enn for tradisjonelle webapplikasjoner. Et AI-prosjekt bestar av mange sammenkoblede komponenter: modelldeployeringer, datasett, pipeline-konfigurasjoner, agentdefinisjoner, tilkoblinger til eksterne tjenester, og tilhoerende infrastruktur som Azure Cosmos DB, Azure AI Search og Azure Storage. Tap av en enkelt komponent kan gjore hele AI-loesningen uoperativ. Saerlig for Foundry Agent Service er tilstandsdata (samtalehistorikk, agent-definisjoner, trad-kontekst) fordelt pa tvers av flere lagringstjenester, og det finnes per i dag ingen innebygd en-klikks eksport/import-funksjon for komplett gjenoppretting. diff --git a/skills/ms-ai-infrastructure/references/bcdr/capacity-planning-dr-configurations.md b/skills/ms-ai-infrastructure/references/bcdr/capacity-planning-dr-configurations.md index f7a489f..fb10062 100644 --- a/skills/ms-ai-infrastructure/references/bcdr/capacity-planning-dr-configurations.md +++ b/skills/ms-ai-infrastructure/references/bcdr/capacity-planning-dr-configurations.md @@ -177,7 +177,7 @@ az cosmosdb sql container throughput update \ 2. **AI Search**: 2 replikaer i stedet for 3 i DR - Besparelse: ~33% på search-kostnaden - - Tradeoff: 99.9% SLA i stedet for 99.99% + - Tradeoff: kun read-only (query) SLA i stedet for full read-write (query + indeksering) SLA — begge er 99,9 % (det finnes ingen 99,99 %-tier) 3. **App Service**: P2v3 i stedet for P3v3, med autoscale - Besparelse: ~50% på compute diff --git a/skills/ms-ai-infrastructure/references/bcdr/network-resilience-patterns-ai.md b/skills/ms-ai-infrastructure/references/bcdr/network-resilience-patterns-ai.md index c1ea0c2..e2ca790 100644 --- a/skills/ms-ai-infrastructure/references/bcdr/network-resilience-patterns-ai.md +++ b/skills/ms-ai-infrastructure/references/bcdr/network-resilience-patterns-ai.md @@ -24,7 +24,7 @@ For norsk offentlig sektor er nettverkssikkerhet regulert gjennom NSMs grunnprin └──────┬──────┘ │ ┌──────▼──────────────────────┐ -│ Azure Front Door (Global) │ ← DDoS Protection Standard +│ Azure Front Door (Global) │ ← innebygd DDoS-beskyttelse (edge) └──────┬──────────────────────┘ │ ┌────┴────┐ @@ -347,7 +347,7 @@ az network vnet peering create \ ### Azure DDoS Protection ```bash -# Aktiver DDoS Protection Standard +# Aktiver DDoS Network Protection az network ddos-protection create \ --name "ddos-ai-protection" \ --resource-group "rg-networking" \ diff --git a/skills/ms-ai-security/references/cost-optimization/azure-ai-foundry-cost-governance.md b/skills/ms-ai-security/references/cost-optimization/azure-ai-foundry-cost-governance.md index 84239cc..85341d7 100644 --- a/skills/ms-ai-security/references/cost-optimization/azure-ai-foundry-cost-governance.md +++ b/skills/ms-ai-security/references/cost-optimization/azure-ai-foundry-cost-governance.md @@ -22,7 +22,7 @@ Quotas er tekniske grenser som kontrollerer hvor mye av en gitt ressurs en subsc | Quota Type | Scope | Default Limit | Adjustable? | |------------|-------|---------------|-------------| -| **Model Quota (TPM)** | Per subscription, per region, per model | Varies by tier (150K-30M TPM) | Yes, via quota request | +| **Model Quota (TPM)** | Per subscription, per region, per model | Varierer per Quota Tier (Free Tier (Tier 0) + Tier 1–6); ca. 100K–225M TPM | Yes, via quota request | | **VM Family Quota** | Per subscription, per region | 24-300 cores (depends on subscription type) | Yes, via support request | | **Compute Instances** | Per region | 500 total compute limit | Yes, up to 2500 via quota UI, beyond via support | | **Serverless API Quota** | Per deployment | 200K TPM, 1K RPM | Yes, one deployment per model per project by default | diff --git a/skills/ms-ai-security/references/cost-optimization/model-selection-price-performance.md b/skills/ms-ai-security/references/cost-optimization/model-selection-price-performance.md index b251a04..46a0689 100644 --- a/skills/ms-ai-security/references/cost-optimization/model-selection-price-performance.md +++ b/skills/ms-ai-security/references/cost-optimization/model-selection-price-performance.md @@ -299,7 +299,7 @@ Månedlig besparelse (NOK) = (Kostnad_current - Kostnad_new) × USD_to_NOK ### Power Platform AI Builder **Modellvalg i AI Builder:** -- AI Builder bruker **Azure OpenAI GPT-4o-mini** som default for generative oppgaver (per desember 2024) +- AI Builder / prompt builder bruker **GPT-4.1 mini** som standardmodell for generative oppgaver (GPT-4o mini og GPT-4o brukes nå kun i US government-regioner) - Ingen direkte modellvalg tilgjengelig i AI Builder-grensesnittet - Kostnader inkludert i AI Builder credits (500 credits/bruker/måned i premium-planer) diff --git a/skills/ms-ai-security/references/cost-optimization/multi-model-strategy-costs.md b/skills/ms-ai-security/references/cost-optimization/multi-model-strategy-costs.md index 8ef78f2..788abf1 100644 --- a/skills/ms-ai-security/references/cost-optimization/multi-model-strategy-costs.md +++ b/skills/ms-ai-security/references/cost-optimization/multi-model-strategy-costs.md @@ -10,7 +10,7 @@ Moderne AI-løsninger krever ofte forskjellige modellkapabiliteter for ulike oppgaver. En multi-model strategy innebærer intelligent routing av requests til den mest kostnadseffektive modellen som tilfredsstiller kvalitetskravene. Med Azure OpenAI-modeller som varierer fra GPT-4.1-nano (59 400 tokens/PTU) til GPT-5 (4 750 tokens/PTU) kan besparelsene være betydelige — opptil 90% kostnadsdifferanse mellom modeller for enkle oppgaver. -Model Router fra Microsoft er en trent språkmodell som automatiserer denne beslutningsprosessen i real-time. Den analyserer prompt-kompleksitet, resonnementskrav og oppgavetype for å velge optimal modell fra et sett på opptil 18 underliggende modeller (inkludert GPT-serien, Claude, DeepSeek, Llama og Grok). Dette gir én deployment-overflate med kombinert kosteffektivitet og kvalitet. +Model Router fra Microsoft er en trent språkmodell som automatiserer denne beslutningsprosessen i real-time. Den analyserer prompt-kompleksitet, resonnementskrav og oppgavetype for å velge optimal modell fra et sett på 28 underliggende modeller (inkludert GPT-serien, Claude, DeepSeek, Llama, Grok og flere; versjon 2025-11-18 oppdateres løpende, så antallet vokser). Dette gir én deployment-overflate med kombinert kosteffektivitet og kvalitet. For organisasjoner som ønsker mer kontroll, tilbyr custom gateway-løsninger (via Azure API Management eller egen kode) mulighet for egendefinerte routing-regler basert på client identity, quota management, blue-green deployments eller data sovereignty-krav. Denne kunnskapsfilen dekker både managed (Model Router) og custom gateway-strategier for multi-model deployments. @@ -24,20 +24,25 @@ For organisasjoner som ønsker mer kontroll, tilbyr custom gateway-løsninger (v | **Routing Modes** | Quality (max nøyaktighet), Balanced (default), Cost (max besparelse) | GA | | **Model Subset** | Custom selection av underliggende modeller for routing | GA | | **Deployment Types** | Global Standard, Data Zone Standard | Regional: East US 2, Sweden Central | -| **Underlying Models** | 18 modeller: GPT-4.1/5-serien, o-series, Claude, DeepSeek, Llama, Grok | Varierer per modell | +| **Underlying Models** | 28 modeller (per 2026-06): GPT-4.1/5/5.2–5.5-serien, o4-mini, Claude (4–5), DeepSeek (V3.1/V3.2), gpt-oss-120b, Llama-4-Maverick, Grok-4 | Varierer per modell | -**Underliggende modeller i Model Router `2025-11-18`:** -- **OpenAI-modeller:** gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-chat, o4-mini, gpt-4o, gpt-4o-mini -- **Reasoning-modeller:** o4-mini (preview) -- **3rd-party modeller:** DeepSeek-V3.1, gpt-oss-120b, Llama-4-Maverick-17B-128E-Instruct-FP8, grok-4, grok-4-fast -- **Claude (krever egen deployment):** claude-haiku-4-5, claude-opus-4-1, claude-sonnet-4-5 +**Underliggende modeller i Model Router `2025-11-18` (28 per 2026-06; oppdateres løpende):** +- **OpenAI:** gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, o4-mini, gpt-5-nano, gpt-5-mini, gpt-5, gpt-5-chat, gpt-5.2, gpt-5.2-chat, gpt-5.3-chat, gpt-5.4-nano, gpt-5.4-mini, gpt-5.4, gpt-5.5 +- **3rd-party:** DeepSeek-V3.1, DeepSeek-V3.2, gpt-oss-120b, Llama-4-Maverick-17B-128E-Instruct-FP8, grok-4, grok-4-fast-reasoning +- **Claude (krever egen deployment):** claude-haiku-4-5, claude-sonnet-4-5, claude-opus-4-1, claude-opus-4-6, claude-opus-4-7 -**Rate limits (Model Router `2025-11-18`):** +**Rate limits (Model Router `2025-11-18`, kvotenivå-basert):** -| Deployment Type | Default RPM | Default TPM | Enterprise RPM | Enterprise TPM | -|-----------------|-------------|-------------|----------------|----------------| -| GlobalStandard | 250 | 250 000 | 400 | 400 000 | -| DataZoneStandard | 150 | 150 000 | 300 | 300 000 | +Grenser skalerer med abonnementets bruksnivå (Quota Tier 1–6), ikke lenger Default/Enterprise. Se [Quota tiers](https://learn.microsoft.com/azure/foundry/openai/quotas-limits). + +| Tier | GlobalStandard RPM | GlobalStandard TPM | DataZoneStandard RPM | DataZoneStandard TPM | +|------|--------------------|--------------------|----------------------|----------------------| +| Tier 1 | 1 000 | 1 000 000 | 300 | 300 000 | +| Tier 2 | 2 000 | 2 000 000 | 670 | 670 000 | +| Tier 3 | 4 000 | 4 000 000 | 1 000 | 1 000 000 | +| Tier 4 | 7 000 | 7 000 000 | 2 000 | 2 000 000 | +| Tier 5 | 10 000 | 10 000 000 | 3 000 | 3 000 000 | +| Tier 6 | 15 000 | 15 000 000 | 4 000 | 4 000 000 | ### Custom Gateway Architectures diff --git a/skills/ms-ai-security/references/cost-optimization/rag-query-cost-reduction.md b/skills/ms-ai-security/references/cost-optimization/rag-query-cost-reduction.md index a5629b9..aa2b4cb 100644 --- a/skills/ms-ai-security/references/cost-optimization/rag-query-cost-reduction.md +++ b/skills/ms-ai-security/references/cost-optimization/rag-query-cost-reduction.md @@ -46,10 +46,10 @@ Basert på Microsoft Learn-data for standard konfigurasjon (5 retrieved document | Tier | Partitions | Replicas | QPS Capacity | Storage | ~NOK/month | Best For | |------|------------|----------|--------------|---------|------------|----------| -| **Basic** | 1 | 3 | Moderate | 2 GB | 1 200 | Proof-of-concept, lav trafikk | -| **S1** | 12 | 12 | High | 25 GB/partition | 2 800 | Produksjon, moderate volumer | -| **S2** | 12 | 12 | Very High | 100 GB/partition | 11 200 | High-volume produksjon | -| **S3** | 12 | 12 | Enterprise | 200 GB/partition | 22 400 | Enterprise-skala | +| **Basic** | 1 | 3 | Moderate | 15 GB/partition (eldre: 2 GB) | 1 200 | Proof-of-concept, lav trafikk | +| **S1** | 12 | 12 | High | 160 GB/partition | 2 800 | Produksjon, moderate volumer | +| **S2** | 12 | 12 | Very High | 512 GB/partition | 11 200 | High-volume produksjon | +| **S3** | 12 | 12 | Enterprise | 1 024 GB/partition | 22 400 | Enterprise-skala | **Baseline (Modellkunnskap):** Prisene er omregnet fra USD til NOK (1 USD ≈ 11 NOK, februar 2026) og er veiledende. diff --git a/skills/ms-ai-security/references/cost-optimization/reserved-capacity-planning.md b/skills/ms-ai-security/references/cost-optimization/reserved-capacity-planning.md index fc329d7..15c4d0b 100644 --- a/skills/ms-ai-security/references/cost-optimization/reserved-capacity-planning.md +++ b/skills/ms-ai-security/references/cost-optimization/reserved-capacity-planning.md @@ -282,7 +282,7 @@ Er tjenesten Azure OpenAI/Foundry Models? **💡 Best Practice:** **ALLTID deploy først, kjøp reservasjon etterpå.** Quota ≠ capacity. -**Autorenew er nå ON som standard for alle nye reservasjoner** (gjelder reservasjoner opprettet etter 2025-Q4). Sjekk innstillingen ved kjøp og deaktiver manuelt hvis ønskelig. Verified (MCP 2026-04). +**Autorenew er opt-in — ikke på som standard.** Aktiveres i Renewal settings eller ved kjøp. Ved utløp kan reservasjonen enten fornyes på samme reservasjons-ordre-ID eller erstattes av en ny reservasjon; en erstatningsreservasjon har autorenew **av** som standard. Sjekk og still inn ønsket fornyelse ved kjøp/fornyelse. Verified (MCP 2026-06). --- diff --git a/skills/ms-ai-security/references/cost-optimization/vector-storage-cost-optimization.md b/skills/ms-ai-security/references/cost-optimization/vector-storage-cost-optimization.md index ef83416..86aa61d 100644 --- a/skills/ms-ai-security/references/cost-optimization/vector-storage-cost-optimization.md +++ b/skills/ms-ai-security/references/cost-optimization/vector-storage-cost-optimization.md @@ -396,10 +396,10 @@ Eksempel: 5 millioner dokumenter, gjennomsnittlig 2000 tokens per dokument | Tier | Vector quota | Pris/måned (1 partition) | Egnet datasett | |------|--------------|--------------------------|----------------| -| **Basic** | 1 GB | ~$75 USD | <100K docs | -| **S1** | 12 GB | ~$250 USD | 100K-1M docs | -| **S2** | 36 GB | ~$1,000 USD | 1M-5M docs | -| **S3** | 72 GB | ~$2,000 USD | 5M-20M docs | +| **Basic** | 5 GB | ~$75 USD | <100K docs | +| **S1** | 35 GB | ~$250 USD | 100K-1M docs | +| **S2** | 150 GB | ~$1,000 USD | 1M-5M docs | +| **S3** | 300 GB | ~$2,000 USD | 5M-20M docs | **Viktig:** Eldre services (pre-April 2024) har lavere quotas. Sjekk oppgraderingsmulighet: `az search service show --name --resource-group `.