

We've all done RAG, now what?
170 snips Sep 29, 2025
Rajiv Shah, Chief Evangelist at Contextual AI and an active content creator, shares insights from his year of building retrieval-augmented generation (RAG) pipelines. He discusses why many AI pilots fail and emphasizes the importance of evaluation and error analysis. Shah explains how RAG differs from traditional model fine-tuning and sheds light on multi-step reasoning in models. He warns against using large language models for every problem and highlights the need for user-focused AI projects that provide measurable value.
AI Snips
Chapters
Transcript
Episode notes
RAG Replaces Retraining For Domain Knowledge
- Retrieval-augmented generation (RAG) finds and supplies relevant knowledge instead of retraining models.
- RAG is widely used for internal knowledge search, HR docs, and customer support.
AI Systems Are Multi-Component Not Single Models
- ChatGPT-like systems are composed of multiple parts such as retrieval, tools, and memory rather than a single per-user model.
- Context engineering manages inputs like retrieval and memory to shape model behavior.
Prioritize Use Cases And Workflow Integration
- Focus on problems and end users, not just shiny new models or demos.
- Ensure solutions integrate into daily workflows and have stakeholder buy-in for adoption.