7.4 End-to-End Solution Architectures
Key Takeaways
- The document processing pipeline combines Document Intelligence + Language + Search: ingest → extract → enrich → index → search.
- The intelligent chatbot architecture combines CLU + Question Answering + Azure OpenAI + Bot Service for multi-capability conversational AI.
- The content moderation pipeline combines Content Safety + OpenAI content filters + human review for comprehensive safety.
- Multi-modal solutions combine Vision + Language + Speech services to process images, text, and audio in a single pipeline.
- Production architectures require logging, monitoring, error handling, retry logic, and graceful degradation.
End-to-End Solution Architectures
Quick Answer: Common AI-102 architectures: document processing (Doc Intel + Language + Search), intelligent chatbot (CLU + Q&A + OpenAI + Bot Service), content moderation (Content Safety + OpenAI filters + human review), and RAG (Search + OpenAI + Content Safety).
Architecture 1: Document Processing Pipeline
[Document Upload]
→ [Azure Blob Storage]
→ [Azure AI Document Intelligence]
├── Extract text, tables, key-value pairs
├── Classify document type (composed model)
└── Extract domain-specific fields
→ [Azure AI Language]
├── Named Entity Recognition
├── Key Phrase Extraction
├── PII Detection and Redaction
└── Language Detection
→ [Azure AI Search]
├── Index enriched content
├── Vector embeddings (Azure OpenAI)
└── Knowledge Store (for analytics)
→ [Power BI / Application]
Use cases: Invoice processing, contract analysis, medical record digitization, compliance document review.
Architecture 2: Intelligent Chatbot
[User Message]
→ [Azure Bot Service] (channel: Teams, Web Chat, etc.)
→ [CLU Orchestration Model]
├── Intent: FAQ → [Custom Question Answering] → Direct answer
├── Intent: Action → [CLU Domain Model] → Extract entities → Call backend API
├── Intent: OpenEnded → [Azure OpenAI + RAG] → Generate grounded response
└── Intent: None → "I don't understand, please rephrase"
→ [Content Safety] → Filter response
→ [Bot Service] → Send response to user
Use cases: Customer support, internal IT helpdesk, product assistant, onboarding guide.
Architecture 3: Content Moderation Pipeline
[User-Generated Content]
→ [Pre-screening]
├── [Azure AI Content Safety] → Text/Image moderation (4 categories)
├── [Blocklist Check] → Organization-specific terms
└── [PII Detection] → Redact personal information
→ [Decision Engine]
├── Safe (severity < 2) → Auto-approve → Publish
├── Borderline (severity 2-3) → Human review queue
└── Harmful (severity >= 4) → Auto-reject → Notify user
→ [For AI-Generated Content]
├── [Prompt Shields] → Block jailbreaks before generation
├── [Azure OpenAI Content Filters] → Screen input + output
├── [Groundedness Detection] → Verify factual accuracy
└── [Protected Material Detection] → Prevent copyright issues
Architecture 4: Enterprise RAG System
[Data Ingestion]
├── [Azure Blob Storage] → Company documents
├── [SharePoint] → Internal knowledge base
└── [Azure SQL] → Structured data
↓
[Azure AI Search Indexer]
├── [Skillset] → OCR + NER + Key Phrases + Embeddings
├── [Vector Index] → Hybrid search (keyword + vector)
└── [Semantic Ranking] → Re-rank by relevance
↓
[Query Pipeline]
├── [User Query] → [Prompt Shields] → Block attacks
├── [Embed Query] → [Vector Search] → Retrieve context
├── [Construct Prompt] → System message + context + query
├── [Azure OpenAI GPT-4o] → Generate response
├── [Content Safety] → Filter harmful output
├── [Groundedness Check] → Verify accuracy
└── [Return Response with Citations]
Production Deployment Checklist
| Category | Requirement | Implementation |
|---|---|---|
| Security | No API keys in code | Managed identity + Key Vault |
| Security | Network isolation | Private endpoints + VNet |
| Security | Access control | RBAC with least privilege |
| Reliability | High availability | Multi-region deployment |
| Reliability | Error handling | Retry with exponential backoff |
| Reliability | Graceful degradation | Fallback responses when AI fails |
| Monitoring | API metrics | Azure Monitor diagnostics |
| Monitoring | Content safety | Log filter triggers |
| Monitoring | Cost tracking | Azure Cost Management alerts |
| Compliance | Data privacy | Encryption, data residency, no model training |
| Compliance | Responsible AI | Content filters, human oversight, transparency |
On the Exam: Architecture questions test whether you can combine multiple Azure AI services correctly. Focus on understanding which service handles which step and the correct order of operations (e.g., content safety checks happen BEFORE the response reaches the user, not after).
In an intelligent chatbot architecture, the CLU orchestration model classifies a user message as an FAQ question. Where should the message be routed?
In a RAG pipeline, what is the correct order of operations after a user submits a query?
A user submits a document that could be an invoice, receipt, or purchase order. Which Document Intelligence feature handles routing automatically?