Handbook
LAUNCHING NOV 9
The RAG Production Handbook
Ship a RAG system that doesn't fall over. ~95 pages on production deployment, cost modelling, observability, and real case studies — with a working code repo.
Launching November 9, 2026.
Subscribe for a free preview chapter + first-20 pricing at ₹699 (regular launch price ₹999).
Launch pricing ends Dec 1 — price rises to ₹1499 after.
Who this is for
Engineers who have built a RAG tutorial that works on a laptop and now need to ship one that handles real traffic. Backend, platform, and AI engineers with 1-5 years of experience.
Not the target: pure researchers, or anyone who hasn't built a RAG prototype yet (start with the free pillar).
What's inside — 10 chapters, 5 appendices, ~95 pages
- Chapter 1 — RAG in one page. The 4-step loop, when RAG is the right tool, when something else is cheaper.
- Chapter 2 — Chunking that doesn't lose information. Decision table across 5 strategies. Plus the edge cases: multilingual, PDFs with tables, legal contracts, chunker versioning, delta reindex pattern that cuts embedding cost 99% on small edits.
- Chapter 3 — Vector database selection in production. Metadata-DB + vector-DB separation pattern. HNSW tuning per vendor. Backup + disaster recovery per vendor. 4 multi-tenancy patterns.
- Chapter 4 — Evaluation that keeps the system honest. 4 tested LLM-as-judge prompts. CI harness in pytest + GitHub Actions. Statistical significance (Wilcoxon). 5-step human review protocol for building a 50-query test set.
- Chapter 5 — Production deployment recipes. Three complete recipes with runnable configs: Vercel + managed vector DB; Fly.io/Railway + pgvector; Kubernetes + self-hosted Qdrant. Each with cost tables across 1K/10K/100K queries.
- Chapter 6 — Cost calculator for RAG at scale. The E+R+G+C per-query formula. Per-vendor pricing shape. Prompt caching, model routing, retrieval-cost optimization. Spreadsheet in Appendix C.
- Chapter 7 — Observability for RAG. 7-field query log schema. 3 alerts worth your pager + 10 that aren't. 6-panel Grafana dashboard. The debug script that solves "the answer was wrong" complaints in 5 minutes.
- Chapter 8 — Three real case studies. B2B support bot (~50 queries/day), internal docs search (~200 queries/day), public dev Q&A (~2500 queries/day). Full architectures, what broke, costs at steady state.
- Chapter 9 — Vendor matrix + decision tree. 6 vendors (adds Weaviate + LanceDB). Index-tuning cheat sheet per vendor. Honest benchmark reality check.
- Chapter 10 — Migration recipes. Full scripts for Chroma → pgvector, Pinecone → Qdrant, pgvector IVFFlat → HNSW. Dual-write pattern for zero-downtime cutover.
Plus — the production code repo
Not just descriptions. A working, forkable reference implementation that handles the ops concerns the toy doesn't:
chunkers.py— all 10 chunking strategies as a drop-in moduleretriever.py— abstracts over pgvector, Qdrant, Pinecone with one.search()callevaluator.py— the test harness from Chapter 4logger.py— the 7-field production logging schemaDockerfile+docker-compose.yml— local dev setup- Full deployment configs for Vercel, Fly.io, and Kubernetes
What this handbook deliberately does NOT cover
- Pure NLP research or embedding model training (different book)
- Multimodal RAG (image + text) — too niche for v1
- Agentic workflows / RAG + tool use — different mental model
- Fine-tuning instead of RAG — covered briefly in the "when to skip" article, not product-level
Format + delivery
- Format: PDF handbook (~95 pages) + production code repo
- Delivery: files emailed within seconds of payment
- Launch price: ₹999 (first 20 buyers get ₹699 in exchange for honest testimonial). Price rises to ₹1499 on Dec 1, 2026.
- License: Single-buyer. Use in your own projects (personal or commercial). Do not redistribute, resell, or share the handbook itself.
- Updates: Free updates for 1 year as vendor APIs + pricing drift
- Support: hello@aiunplugged.in — response within 24 hours
FAQ
How is this different from the free RAG articles on aiunplugged?
The free cluster gets you from zero to "RAG tutorial that works on my laptop." The handbook gets you from that to "RAG system in production that doesn't fall over at 3am." Chapters 1-4 compress the free material into tighter production form; Chapters 5-10 are entirely new paid-only content (deployment, cost, observability, case studies, vendor migrations) plus the working code repo.
Which LLM / embedding model / vector DB does it assume?
None, specifically. The handbook is vendor-agnostic: Claude, OpenAI, local models for generation; OpenAI, Cohere, sentence-transformers for embeddings; pgvector, Qdrant, Pinecone, Chroma, Weaviate, LanceDB for vector storage. Each section says when each choice wins and when it doesn't.
Will this work for a non-English corpus?
Yes — Chapter 2 has a dedicated multilingual section covering tokenizer mismatch, mixed-script retrieval, RTL text, and per-language chunking. The model + embedding recommendations call out multilingual options throughout.
Can I use the code repo in my commercial project?
Yes. The code repo is yours to fork, modify, and ship in any project — personal or commercial — no attribution required. You just can't redistribute the repo itself or the handbook PDF.
Why ₹999 at launch — and why does it rise on Dec 1?
India-first launch pricing. The audience this is written for includes Indian engineers who shouldn't pay USD-equivalent prices for production patterns. After the launch window (Nov 9 - Nov 30), the regular price is ₹1499 — still below comparable technical books ($45 for Designing Data-Intensive Applications, $59+ for Educative System Design). First 20 buyers get ₹699 in exchange for a short honest testimonial.
What if it's not useful?
7-day refund, no questions asked. Reply to the delivery email with "refund please" and the money is back on your card.
Launches November 9, 2026
Subscribe for a free preview chapter + first-20 pricing at ₹699. Launch pricing ends Dec 1 — then ₹1499.