
Building AI Applications: Techniques

AI Engineering with Chip Huyen
Topic 6 of 14
Published Feb 5, 2025 · Clip 18:27–28:39 (10:13)
Chip lays out a development path for AI applications: first define what a good response means (using LinkedIn's job-fit assessment example), then prompt engineering with examples, then evaluation, then RAG. She cautions against jumping straight to vector databases — simple keyword retrieval often works, chunking introduces problems solved by metadata and Anthropic's contextual retrieval, and bm25 is a hard-to-beat baseline. Fine-tuning is a last resort with hosting and maintenance burdens, since new models quickly outpace fine-tuned ones.