Notes & thoughts.

Evals show that a RAG system failed. Traces show whether retrieval, context construction, generation, or the eval itself caused it.

Open-weights models, proprietary models, parameters, distillation, quantization, inference, deployment, and cost without the usual fog.

A compact map of NLP, NLU, NLG, tokenization, entity extraction, and sentiment before embeddings enter the picture.

How text representation moved from one-hot vectors, Bag of Words, and TF-IDF toward dense embeddings that capture meaning-like similarity.

Why static embeddings break on context, how self-attention updates token meaning, and why Transformers changed NLP.

Natural language is not executable. Intent classification turns messy requests into structured actions, metadata, validation, and workflow routing.

A compact map of how retrieval moved from exact keyword matching to agents that decide what to search, inspect, and verify.

Two related ideas that get fused in production code. A simple separation that makes agents easier to debug, eval, and scale.