You followed the tutorials. You embedded your docs, wired up a vector database, and shipped a shiny RAG chatbot...only to watch it confidently invent facts. The standard playbook is letting you down.

The twist? The wins in production are often simple and counter-intuitive. The secret isn't a bigger model, it's smarter retrieval.

This talk reveals the overlooked patterns that turn a POC into a production-ready system:

Surprise #1: Your chunks are wrong.
Stop naïve text splitting. Use your content's native structure (headers, JSON) to keep context intact in your embeddings.

Surprise #2: Keywords still matter.
Hybrid dense+sparse (BM25) raises coverage and precision, especially for entities, IDs, and exact phrases.

Surprise #3: Retrieval is two-stage.
Add reranking with a stronger cross-encoder to sift the true signal from top-k noise.

Surprise #4: You can’t steer blind.
Run evals (e.g. Recall@k) so iterations move the numbers, not just the "vibes."

No RAG 101. We'll show before/after metrics, failure modes, and the trade-offs that separate prototypes from production systems. You'll leave with a pragmatic playbook for retrieval that actually remembers.

Attendees leave with a concrete playbook for high‑fidelity RAG retrieval.

I'll (try to) teach developers how to design smarter retrieval pipelines and prove improvements with metrics. The focus is on structure‑aware chunking, hybrid dense+sparse search, two‑stage retrieval with reranking, and evals that guide iteration beyond vibes.