RAG in Production

• Chunking strategies that actually survive real corpora
• Pick the right retrieval pattern: vector, BM25, hybrid, or rerank
• Build a retrieval eval set — and the metrics that matter
• Handle freshness, deletions, and multi-tenant isolation
• Cut RAG cost 5–10× without losing answer quality

1. Ingestion + chunking

   Session 1

   Token-aware chunking, structural chunking, hybrid splits. When to recompute.

2. Retrieval patterns

   Session 1

   Vector, BM25, hybrid, rerank. Choosing per query class.

3. Eval + drift

   Session 2

   Build a retrieval-eval golden set, detect drift, and ship updates without regressions.

4. Cost + ops

   Session 2

   Caching, prefiltering, cheaper embeddings, and tenant isolation.

• Comfortable with Python
• You have or are planning a RAG system
• Familiar with at least one vector DB

After this workshop you leave with a concrete artefact you built live and a playbook you can use the next week. Cohort chat stays open so you can ask follow-up questions while you ship.

Which vector DB do you teach?+

Patterns generalize. We demo with pgvector and one managed option, but the takeaways apply to any store.

Will you cover graph RAG / agentic retrieval?+

Briefly, with the honest tradeoffs. Most teams should ship hybrid + rerank before reaching for graph.