Mental Model — What This Stack Is and Why It Exists
Microsoft assembled the Azure AI ecosystem to let enterprises build production-grade AI without managing infrastructure. The three services in this JD form a natural RAG pipeline: Document Intelligence extracts structured data from raw documents (PDFs, forms, handwritten notes); Azure AI Search indexes it with vector + keyword hybrid retrieval; Azure OpenAI generates answers grounded in those retrieved chunks. Think of it as: extract → index → query → generate. Your existing work maps directly — you've built all four phases, just on different platforms.
Document Intelligence (formerly Form Recognizer)
- OCR + ML extraction from PDFs, images, forms
- Prebuilt models: invoices, receipts, ID cards, health insurance cards
- Custom models: train on your own document layouts
- Outputs: structured JSON (fields, values, bounding boxes, confidence scores)
- AF use case: Extract structured data from arbitration demand letters, arbitration awards, subrogation claim forms
Your bridge: "At NewsRx I built unstructured → structured pipelines for 950+ biomedical PDFs; Document Intelligence is Azure's managed version of that extraction layer."
Azure AI Search (formerly Cognitive Search)
- Enterprise search + vector store + hybrid retrieval engine
- Supports dense vector search (embeddings) + BM25 keyword search, merged via RRF
- Semantic ranker: cross-encoder reranking for relevance
- Integrated with Azure OpenAI for RAG pipelines natively
- Skillsets: built-in NLP enrichments (NER, key phrase, sentiment, OCR) in the indexer pipeline
- AF use case: Index all historical arbitration records; retrieve relevant precedents for an incoming dispute
Your bridge: "Architecturally equivalent to Elasticsearch with Azure-native embedding and reranking. I've built vector retrieval pipelines; this is the managed Azure surface for that."
Azure OpenAI Service
- GPT-4o, GPT-4 Turbo, GPT-3.5 via Microsoft-hosted Azure endpoints
- Same models as OpenAI — private deployment, data stays in your Azure tenant
- API-compatible with OpenAI SDK (just swap endpoint + key)
- Supports: chat completions, embeddings (text-embedding-3-small/large), function calling, structured outputs
- Used with Azure AI Search for RAG: retrieve → stuff context → generate answer
- AF use case: Generate arbitration summaries, classify dispute types, answer queries against historical awards
Your bridge: "I've used Anthropic Claude API for production GenAI at NewsRx and Amgen. Azure OpenAI is the same call pattern — swap the endpoint. The architecture is identical."
The RAG Pipeline This Stack Builds — End to End
Raw Documents
(PDFs, emails,
forms)
Document
Intelligence
Extract + OCR
Azure AI Search
Index (vector +
keyword)
Query + Retrieve
Top-K chunks
(hybrid RRF)
Azure OpenAI
Generate grounded
answer
Each stage maps to your existing work: NewsRx bronze→silver→gold pipeline = Document Intelligence + Search indexer. Invistics multi-EHR normalization = ingestion + extraction. Amgen multi-agent GenAI = Azure OpenAI orchestration layer.
Key Concepts — 8 Terms to Own
| Term | One-sentence definition |
|---|
| Hybrid Retrieval | Combining dense vector search (semantic similarity) with sparse BM25 keyword search; merged via Reciprocal Rank Fusion (RRF) for better recall than either alone. |
| Semantic Ranker | Azure AI Search's cross-encoder reranking layer that rescores retrieved results for contextual relevance beyond keyword/vector match. |
| Skillset / Indexer | Azure AI Search pipeline that enriches documents during indexing — runs OCR, NER, key phrases, sentiment as pre-processing steps before documents hit the index. |
| Embedding | Vector representation of text in high-dimensional space; similar meaning → similar vectors; used for vector search in Azure AI Search. |
| Chunking Strategy | How documents are split before embedding; affects retrieval quality — fixed-size, sentence-based, or semantic chunking; overlap prevents context loss at boundaries. |
| Grounding | Constraining LLM generation to retrieved context only; prevents hallucination by instructing the model to answer only from provided chunks. |
| Custom Extraction Model | Document Intelligence model trained on specific form layouts; used when prebuilt models don't match your document structure. |
| Managed Identity / RBAC | Azure's way of granting services access to each other without storing credentials; critical for PII/PHI compliance in production. |
Your Validation Angle — QA Lens on This Stack
Extraction Validation (Document Intelligence)
- Confidence score thresholds: flag low-confidence extractions for human review
- Ground truth comparison: validate model output against manually extracted sample
- Field-level accuracy: track per-field extraction accuracy; some fields matter more (amounts vs. dates)
- Your precedent: At Amgen, all AI outputs required human signatory review; same pattern applies here
Retrieval Validation (Azure AI Search)
- Precision@K and recall@K on test query sets
- Relevance judging: SME-rated query-result pairs for semantic ranker calibration
- Chunk quality: too large (context dilution) vs. too small (context loss) — evaluate with downstream answer quality
- Your precedent: At Invistics, feature validation used domain expert panel judgment alongside ML metrics — same approach for retrieval quality
Interview Language — 5 Phrases to Use
"The RAG pipeline I'd propose follows the Azure-native pattern — Document Intelligence handles extraction, AI Search manages the hybrid retrieval index, and Azure OpenAI generates grounded responses with citation back to source documents."
"I've implemented production RAG in a regulated environment at Amgen — the grounding discipline and hallucination mitigation patterns I built there translate directly to this Azure stack."
"For document extraction at scale, I'd start with Document Intelligence prebuilt models on the arbitration form types, then train custom models on any layouts that don't reach confidence threshold."
"Hybrid retrieval — vector plus BM25 merged via RRF — consistently outperforms either alone for domain-specific corpora; that's what I'd configure in Azure AI Search from day one."
"PHI/PII masking in this stack sits at the indexer skillset level — you can filter or redact before content reaches the LLM, which is the right governance pattern for insurance claim data."