Speak at The Fifth Elephant 2026 Annual Conference

Almost everyone’s RAG journey started the same way: chunk the documents, embed them, retrieve the top-k, stuff the prompt, generate. Then reality hits — semantic drift returns wrong-but-similar chunks, multi-hop questions fall apart because each chunk was embedded in isolation, “summarize across everything” has no good answer, and every attempt to keep data fresh inflates latency. The field has quietly moved through several generations of answers to these failures, but most teams reach for whatever’s trending — GraphRAG here, “just add more context” there — without a clear picture of what each approach actually fixes or what it costs.

A practitioner’s map of how RAG has evolved across four generations — naive/semantic retrieval, structured RAG (graphs, trees, PageRank-style ranking), reconciled-memory and world-model systems, and the 2026 frontier of self-correcting neuro-symbolic hybrids, agent-curated knowledge graphs, knowledge orchestration beyond triples, and parametric knowledge injection that pushes facts into the model’s weights. We’ll read every generation through the same three axes — how knowledge is represented, where the “intelligence” runs, and what it costs — so the through-line is obvious: each generation buys better reasoning, then claws back the cost the previous one introduced. It’s drawn from recent research and informed by building and operating a system that runs several of these architectures side by side.

1–2 takeaways from your session

• A reusable mental model — three axes (representation × where intelligence runs × cost) and four generations — to evaluate any RAG approach and match it to your workload instead of chasing hype.
• A grounded tour of the 2026 frontier (self-correcting vector+graph hybrids, multi-agent graph curation, agent-native knowledge orchestration, and parametric injection) and the concrete quality-vs-cost trade-off each one makes.

Which audiences is your session going to be beneficial for?

AI/ML and data engineers building or scaling retrieval and knowledge systems; architects and tech leads choosing a RAG strategy; Useful for both hands-on builders and decision-makers. No prior knowledge beyond basic RAG (chunk → embed → retrieve) is assumed.

Draft slides

https://docs.google.com/presentation/d/1-4iv1tZk1SoC3k2QQ6GNIbjt17lnJr1J/

Bio

ML Engineer at Nutanix

https://www.linkedin.com/in/shivam96/
https://medium.com/@shivu-agr