01 · 16 min · 2026-09-15
RAG Architecture in Production: Building a 12-Component System
A deep technical guide to production RAG with hybrid search, pgvector, temporal intelligence, and enterprise-grade chunking strategies.
Retrieval-augmented generation as it behaves in production: choosing an architecture and a store, keeping tenants apart, and surviving a compliance review.
9 guides · updated 13 September 2026
01 · 16 min · 2026-09-15
A deep technical guide to production RAG with hybrid search, pgvector, temporal intelligence, and enterprise-grade chunking strategies.
02 · 14 min · 2026-08-28
Hybrid RAG, GraphRAG, Agentic RAG, Corrective RAG and Multimodal RAG compared on retrieval quality, reasoning, reliability and cost - with a decision guide.
03 · 13 min · 2026-08-28
pgvector, Pinecone, Qdrant, Weaviate and Milvus compared on the things that decide production RAG: hybrid search and metadata filtering, not benchmark speed.
04 · 9 min · 2026-03-20
RAG for knowledge retrieval, fine-tuning for behaviour change. A practical decision framework with cost comparisons from real projects.
05 · 13 min · 2026-09-09
A metadata filter is not tenant isolation. The four independent layers that keep one customer from retrieving another, from a live multi-tenant RAG system.
06 · 13 min · 2026-09-09
Your retrieved documents are untrusted input. What actually reduces prompt injection risk in a production RAG pipeline, and what only looks like it does.
07 · 13 min · 2026-09-09
Deleting a document does not delete its embeddings, and soft delete is not erasure. How to make Article 17 work across a RAG pipeline.
08 · 12 min · 2026-08-11
How to choose a RAG development partner: the nine components a real quote covers, what SOC 2 and HIPAA change, and seven questions that expose a demo-builder.
09 · 12 min · 2026-09-09
RAG development partners compared by category, with the selection criteria published. Boutiques, dev shops, consultancies and independents - what each is actually good at.
These write-ups come out of production systems rather than reading. If you are building in this area, here is how I work on it — and the benchmark results are the numbers behind the claims, including the ones that look bad.