RAG end-to-end¶

The second recipe on top of Foundations and Call an LLM. Same tri-parity discipline, one more step: retrieval-augmented generation.

The point of E3 is that the pieces compose. The E2 chat service is the generation step here — E3 does not rewrite it. Everything the reader already learned about Azure OpenAI / Vertex Gemini / Bedrock stays intact; RAG just wraps retrieve around generate.

The pipeline¶

     ┌──────────┐   ┌──────────┐   ┌────────────┐   ┌──────────────┐
     │  embed   │──►│  store   │──►│  retrieve  │──►│  generate    │
     │ (model)  │   │ (vector) │   │  (top-K)   │   │  (E2 chat)   │
     └──────────┘   └──────────┘   └────────────┘   └──────────────┘
       per-cloud   dual-store       dual-store         REUSED verbatim
       flagship    (native + pgvector)                 from Call an LLM

Two storage paths per cloud¶

Every provider gets both storage backends. Pick whichever fits.

Cloud	Native vector store	Provider-agnostic
Azure	Azure AI Search (`azure-search-documents`)	pgvector on Azure Database for PostgreSQL Flexible Server
GCP	Vertex AI Vector Search	pgvector on Cloud SQL for PostgreSQL
AWS	Bedrock Knowledge Bases (`bedrock-agent-runtime`)	pgvector on RDS for PostgreSQL

The native side is where you sign up for the cloud's fully-managed features (semantic ranking, hybrid, reranking). The pgvector side is where you own the store — same Python, same SQL, same schema on all three clouds. Pick the row where the trade-off lands right for your team.

Docs verified 2026-08-08¶

Every model id + SDK below was verified against these live pages:

Azure OpenAI embeddings (text-embedding-3-small, text-embedding-3-large) — learn.microsoft.com/…/openai/how-to/embeddings
Azure AI Search vector quickstart (azure-search-documents, VectorizedQuery) — learn.microsoft.com/…/search/search-get-started-vector
Vertex AI embeddings (gemini-embedding-001, text-embedding-005) — ai.google.dev/gemini-api/docs/embeddings
Bedrock Titan embeddings (amazon.titan-embed-text-v2:0, 1024 dims) — docs.aws.amazon.com/bedrock/…/titan-embedding-models
Bedrock Knowledge Bases retrieve API (bedrock-agent-runtime, retrieve_and_generate) — docs.aws.amazon.com/bedrock/…/kb-test-config
pgvector — github.com/pgvector/pgvector

Pages¶

Embed — the embedding model per provider + a common Python signature the app uses.
Store — native — Azure AI Search / Vertex Vector Search / Bedrock KB.
Store — pgvector — one SQL schema, three managed Postgres flavors.
Retrieve — embed the query, top-K similarity.
Generate — feed retrieved chunks into the E2 chat service. This is a reuse chapter, not a rewrite.
Deploy — Terraform composing the E1 baseline + provider vector store + Postgres, Helm reusing the E2 chart.

Validate-only¶

Every gate on this batch is a schema/lint check — terraform validate, helm lint, mkdocs build --strict, python -m compileall, plus Chiron's completeness-check + leak-scan. Live LLM + live vector-store calls are laptop operations; the "verify" block on each page shows the one-liner. Phase-3 billed apply lands separately.