Azure AI Studio (ตอนนี้เป็นส่วนหนึ่งของ Azure AI Foundry) เป็นแพลตฟอร์มพัฒนา AI แบบ Pro-Code (code-first) สำหรับ AI developers และ data scientists สร้าง custom generative AI applications (copilots, chatbots, RAG systems, AI agents) ตั้งแต่การทดลอง (experimentation) จนถึง production deployment
ความสามารถหลัก
- Model Catalog: ค้นหาและเลือกโมเดลจาก Azure OpenAI (GPT-4o, GPT-4, DALL-E 3), Open Source (Llama 3, Phi-3, Mistral, Nemotron), และ third-party (Cohere, Jina AI, NVIDIA) - 1,600+ models
- Prompt Flow: Visual authoring tool สำหรับ design, test, evaluate, deploy LLM flows (prompt engineering, chaining, RAG pipelines) - support code-first (Python) และ visual
- Azure AI Search (Vector Search): Integrated vector database สำหรับ RAG (Retrieval-Augmented Generation) - hybrid search (vector + keyword), semantic ranking
- Fine-tuning: ปรับโมเดลสำหรับ domain-specific tasks (GPT-3.5-Turbo, GPT-4o, Llama, Phi, Mistral)
- Evaluation: Automated evaluation (groundedness, relevance, fluency, safety) และ human evaluation workflows
- Content Safety: Built-in content filtering (violence, hate, sexual, self-harm) และ custom blocklists
- Deployment & Monitoring: Managed online endpoints (auto-scaling, authentication, logging) + monitoring (metrics, traces, model drift)
- Responsible AI Dashboard: Fairness, interpretability, error analysis, counterfactuals
- AI Safety Evaluations: Red teaming, jailbreak testing, automated safety benchmarks
วงจรชีวิตการพัฒนา AI ใน Azure AI Studio
- Explore: Browse model catalog, benchmarks, sample prompts
- Build: Prompt flow development, RAG setup, fine-tuning, prompt engineering
- Evaluate: Run evaluations (automated + human) เปรียบเทียบ models/prompts
- Deploy: Push ไป production endpoints (managed online endpoints)
- Monitor: Observability (metrics, logs, traces), model drift detection, data drift
- Govern: Model cards, lineage, compliance, access control (RBAC)
Azure AI Studio vs. Copilot Studio vs. Azure OpenAI Service
| Aspect | Azure AI Studio | Copilot Studio | Azure OpenAI Service |
|---|---|---|---|
| Target Audience | AI Developers, Data Scientists (Pro-code) | Business Users, Makers (Low-code/No-code) | Developers integrating OpenAI models |
| Primary Focus | Custom AI app development (end-to-end) | Conversational AI agents (copilots) | Access to OpenAI models via API |
| Model Support | 1,600+ models (OpenAI, OSS, third-party) | Azure OpenAI models (via integration) | OpenAI models only |
| RAG/Vector Search | Built-in Azure AI Search integration | Via connectors / Dataverse | DIY (bring your own vector DB) |
| Fine-tuning | Supported (multiple models) | Not directly | Supported (GPT-3.5-Turbo, GPT-4o) |
| Evaluation/Safety | Comprehensive (automated + human) | Basic analytics | Content filtering only |
